3D printing

Abstract

A method 500 of generating process parameters for a 3D printing process comprises obtaining 502 vibrational spectroscopic data, e.g. an Infra-Red – IR – spectrum, of a material to be printed; providing 504 the spectroscopic data of the material to be printed to a machine learning model, the machine learning model having been trained on a dataset of spectroscopic data of sample materials to predict at least one of (i) the material of a sample, and (ii) one or more thermophysical properties of a sample from spectroscopic data; predicting 506 at least one of (i) the identity of the material to be printed, and (ii) one or more thermophysical properties of the material to be printed using the machine learning model; and determining 508 one or more process parameters for the printing process using the material to be printed based on the at least one prediction regarding the material to be printed. The method may further comprise controlling a 3D printing process based on the generated process parameters. A 3D printer arranged to perform such a method 500 is also disclosed.

Description

[0001] 3D PRINTING

[0002] The invention relates to three-dimensional (3D) printing, and in particular to characterisation of a material to be printed (the characterisation optionally being performed in-line), optionally with action taken based on that characterisation, and to printing mode -switching options. The disclosure may have particular relevance to multi-model 3D printers, and in particular to 3D printers which can switch between more traditional filament extrusion-based 3D printing and electrohydrodynamic jet printing and spraying 3D printing techniques.

[0003] Current additive manufacturing technologies have an intrinsic antagonism between their degree of structural resolution and their manufacturing speed or “printing efficiency”. Resolution is primarily dictated by the size of the printing nozzle, which directly affects the size of the meniscus of the “ink” formed at the nozzle tip (it will be appreciated that the “ink” for 3D printing may be a polymeric filament, or indeed a solution for solution electrospinning, and that the term “ink” is used here in the general sense of a material to be printed). In an attempt to get around these limitations, the electrohydrodynamic (eHD) emission phenomenon has been coupled with more traditional additive manufacturing to provide a multi-modal, multidimensional printing technology. However, this multi-modal printing technology has not been made commercially viable due to difficulties in control - instability in product quality and the need for careful post-printing analysis of all parts to check for both internal and external defects have held the technology back. There have been difficulties in integrating the differing requirements and limitations of various printing modes into an automated device.

[0004] According to a first aspect of the invention, there is provided a method of generating process parameters for a 3D printing process, the method comprising: obtaining vibrational spectroscopic data of a material to be printed; providing the spectroscopic data of the material to be printed to a machine learning model, the machine learning model having been trained on a dataset of spectroscopic data of sample materials to predict at least one of (i) the material of a sample, and (ii) one or more thermophysical properties of a sample from spectroscopic data; predicting at least one of the identity of the material to be printed, and the one or more thermophysical properties of the material to be printed using the machine learning model; and determining one or more process parameters for the printing process using the material to be printed based on the at least one prediction regarding the material to be printed.

[0005] Embodiments may therefore apply machine learning methods to map the spectroscopic data to printing parameters for 3D printing without human input. Use of the machine learning (ML) model therefore allows automatic / autonomous determination of one or more of the process parameters for 3D printing.

[0006] The vibrational spectroscopic data may be one or more spectra obtained from a vibrational spectroscopic technique, e.g. an Infra Red (IR) spectrum, and optionally a Fourier Transform Infrared - FTIR - spectrum, and / or a near Infrared (NIR) spectrum. Alternatively or additionally, the vibrational spectroscopic data may be obtained by Raman spectroscopy, or another form of vibrational spectroscopy in which the vibrational modes of molecules of the material are observed. It will be appreciated that the same type of spectroscopic data is obtained as the machine learning model was trained on, be that Raman spectra, FTIR spectra, or another kind of vibrational spectroscopic data. In some embodiments, one or more other material properties - e.g. material density - may be included as part of the training data for the machine learning model which is trained to predict at least one of the material of a sample, and one or more thermophysical properties of a sample. These one or more properties of the material to be printed may also be measured (optionally in-line), or otherwise provided to the system implementing the method, and used alongside the spectroscopic data (e.g. an IR spectrum) in predicting the material and / or its thermophysical properties.

[0007] The material to be printed may be a single material, or may be primarily a single material with only traces of other materials. For example, the material may be, or primarily comprise, a polymer such as acrylonitrile butadiene styrene (ABS), polylactic acid (PL A), or poly ether ester ketone (PEEK). In such cases, the identity of the material may be the name and / or chemical structure of the (primary) material. The material to be printed may be a mixture of multiple different substances - the identity of the material may therefore comprise a list of materials making up the material to be printed and an estimation of the percentage of each (by weight or by volume, as desired).

[0008] In some embodiments, one or more predicted thermophysical properties may be provided instead of the identity of the material. In other embodiments, both the material’s identity and one or more thermophysical properties may be predicted.

[0009] The step of determining the one or more process parameters for the printing process may also be performed using a machine learning model, that machine learning model having been trained on an initial dataset correlating polymer chemistry / thermophysical properties and / or specific named polymers / other materials to appropriate process parameters, the initial dataset having been generated from experimental results (optionally manually-generated). The model trained on this dataset can then predict suitable process parameters for other materials, extrapolating beyond the materials covered by the initial dataset. The method may further comprise controlling a 3D printing process based on the generated process parameters.

[0010] The step of obtaining the vibrational spectroscopic data may comprise obtaining data from a spectrometer arranged in-line with the printing process, such that a vibrational spectrum is generated for the material to be printed whilst the material is stored in a store of the 3D printer, or as the material moves towards, or emerges from, a printing nozzle. In some embodiments, the spectrum is therefore taken in-line, before the material to be printed is fed to a nozzle of a 3D printer. For example, a probe may be connected to, or provided as part of, a 3D printer such that vibrational spectra can be collected as the material to be printed is moved towards a printing head of a 3D printer. For the example of using IR spectra, the step of obtaining the IR spectrum may comprise obtaining data from an IR spectrometer arranged in-line with the printing process, such that an IR spectrum is generated for the material to be printed whilst the material is stored in a store of the 3D printer, or as the material moves towards, or emerges from, a printing nozzle. In some embodiments, the IR spectrum is therefore taken in-line, before the material to be printed is fed to a nozzle of a 3D printer. For example, an IR probe may be connected to, or provided as part of, a 3D printer such that spectra can be collected as the material is moved towards a printing head of a 3D printer.

[0011] The application of machine learning methods directly to in-line characterisation so as to make predictions therefore allows for the autonomous selection of processing conditions in situ, providing a “smart” printer that can adapt automatically to a change in the material to be printed. The intelligent technology also provides a characterisation tool for “unknown” polymeric materials in the feed. Printing parameters can therefore be adjusted and set automatically based on the polymeric material(s) in the feed, without a user needing to provide information on - or even know - the type of printing material in use. In other embodiments, one or more vibrational spectra (e.g. an IR spectrum) may be taken of a material before it is provided to a 3D printer (e.g. as an analysis step prior to loading a 3D printer for a printing process) or after the material has been printed (e g immediately after an extrusion has been laid down, or later). It will be appreciated that in embodiments in which the spectrum is taken and predictions made after an extrusion has been laid down, the predictions may then be used to adjust the parameters for the current, ongoing, printing process, and / or to save tailored parameters for a next usage of the same material.

[0012] Where IR spectra are used, the IR spectra may be or comprise Fourier Transform Infrared - FTIR - spectra, and / or may or comprise near Infrared (NIR) spectra.

[0013] The method may further comprise controlling the printing process in accordance with the one or more determined process parameters. Automated adjustments may therefore be made based on knowledge of the printing material, without a user’s input. In various embodiments, the same approach may be used when the material to be printed is (i) a filament for Fused Filament Fabrication (FFF), (ii) a solution for solution eHD, or (iii) a material for melt eHD. It will be appreciated that some process parameters may be relevant to all 3D printing modes, and some to only a subset of 3D printing modes (with current and voltage being generally inapplicable to FFF, for example). The printing process parameters may include one or more of: Layer Height, Print Speed, Extruder / Nozzle Temperature, Bed Temperature, Cooling Fan Speed, Voltage, and Bed- Nozzle Distance. In some embodiments, some or all of the 3D printer may be enclosed within a chamber - in such embodiments Humidity within that chamber, and / or Chamber Temperature may also be controlled. Such a chamber may be particularly beneficial for eHD printing modes, and may therefore be referred to as an eHD chamber.

[0014] The printing process may be a fused filament fabrication - FFF - 3D printing process, or an electrohydrodynamic - eHD - printing process. The, or a, determined process parameter may be a selection of printing mode based on the one or more materials available to it and the desired morphology and target properties of the part to be printed. E.g. FFF may be selected if the FFF filament available is of a suitable material to form a sufficiently fine extrudate for the current part to be printed, and an eHD mode may be selected instead if the FFF filament available could not be reliably used to form a thin enough extrudate.

[0015] In embodiments in which the printing process is a fused filament fabrication - FFF- 3D printing process, the at least one process parameter may be or comprise one or more of:

[0016] (i) filament diameter;

[0017] (ii)extruder temperature / hot-end temperature;

[0018] (iii)feed rate (e.g. filament feeding rate into the melt or FFF nozzle, or the syringe pump rate for the solution ehD);

[0019] (iv)printing speed (i.e. movement speed of printing head relative to the printing bed);

[0020] (v) printing shear force;

[0021] (vi)bed temperature (i.e. temperature of the printing bed);

[0022] (vii) nozzle temperature;

[0023] (viii) raster angle;

[0024] (ix) nozzle angle; and

[0025] (x) nozzle diameter.

[0026] The raster angle, is known to significantly affect mechanical properties of the output, in particular the tensile strength and elastic modulus of printed parts, and potentially also flexural strength and modulus (affecting direct bending). Desired values for one or more of these properties of the final 3D-printed product may therefore be used in conjunction with material identification to determine an optimal raster angle. In some embodiments, nozzle angle and / or nozzle diameter may be adjustable in use for a specific nozzle, e.g. having a controllable, expandable, nozzle. In other embodiments, nozzle angle or nozzle diameter may be changed by switching to a different nozzle, optionally mounted on the same print head. In still other embodiments, nozzle angle and / or nozzle diameter may only be adjustable by removing and replacing a nozzle of the 3D printer - nozzle angle and / or nozzle diameter may or may not be listed amongst the process parameters to be determined in such embodiments.

[0027] It will be appreciated that speed of axial movement of the printing head or nozzle influences shear force, which is strongly dependent on rheology of the material being printed - shear force in printing can therefore be tuned with knowledge of the material’s properties, and control of temperature and printing head or nozzle movement speed.

[0028] In particular, in some embodiments the at least one process parameter may be or comprise one or more of: filament diameter; feed rate; printing speed; printing shear force; bed temperature; and hot-end or nozzle temperature, in various FFF implementations. In embodiments in which the printing process is a material jetting or spraying 3D printing process, and in particular is an eHD printing process, the at least one process parameter may be or comprise one or more of:

[0029] (i) whether to use melt or solution (e.g. looking at the available melt material(s) and the available solution(s) as identified from the spectroscopic data, and the desired properties of the part to be printed);

[0030] (ii) voltage;

[0031] (iii) extruder temperature;

[0032] (iv) nozzle height;

[0033] (v) flowrate of melt or solution;

[0034] (vi) mode of operation (e.g. melt or solution electrospinning, electrospraying);

[0035] (vii) Taylor cone modality (e.g. single cone or MultiJet);

[0036] (viii) temperature of the nozzle;

[0037] (ix) temperature of the eHD chamber (where present);

[0038] (x) humidity of the eHD chamber (where present); and

[0039] (xi) printing speed.

[0040] For a selection between the use of a melt or a solution, key factors to consider include the desired degree of fibre alignment / orientation in the 3D printed product, and the desired fibre size (diameter) in the 3D printed product. Solution electrospinning generally provides finer fibres with arbitrary orientations / little alignment, whereas melt electrospinning produces thicker fibres with better stability (more aligned).

[0041] For more details on jetting regimes and cone modalities, the reader is directed to A. Lee, H. Jin, H - W. Dang, K.-H. Choi, K. H. Ahn, "Optimization of experimental parameters to determine the jetting regimes in electrohydrodynamic printing” , Langmuir 29 (44) (2013) 13630-13639, and R. Hartman, D. Brunner, D. Camelot, J. Marijnissen, B. Scarlett, "Electrohydrodynamic atomization in the cone-jet mode physical modelling of the liquid cone and jet”, Journal of Aerosol Science 30 (7) (1999) 823-849.

[0042] The printing process may be a multi-modal printing process, capable of switching between printing modes (e.g. FFF, melt or solution electrospinning, electrospraying, etc.) whilst using the same 3D printer. In such embodiments, the at least one process parameter may be or comprise a selection of a printing mode. The selection may optionally be made between FFF 3D printing, melt electrospinning, and solution electrospinning. In multi-modal printing processes, the method may further comprise: detecting a current flowing within printing control circuitry of a 3D printer implementing the method, inferring in which printing mode the process is currently operating from the current, and stopping the printing process if there is a mis-match between the inferred printing mode and an intended printing mode. If the current is approximately zero (zero within errors), the mode may be identified as “standard” / ”traditional”, nonelectrohydrodynamic, 3D printing: FFF 3D printing. If the current is non-zero, the mode may be identified as an eHD printing mode. In eHD modes, current is generally proportional to the size of fibres. As solution electrospinning generally produces finer fibres, a lower current may indicate that the 3D printer is operating in a solution electrospinning mode rather than a melt electrospinning mode, with precise limits depending on materials and / or other factors.

[0043] In an eHD printing processes, the method may further comprise: detecting a current flowing within printing control circuitry of a 3D printer implementing the method; inferring fibre size from the current; and stopping or adjusting the printing process if there is a mis-match between the inferred fibre size and an intended fibre size.

[0044] The method may be arranged to predict at least one of the glass transition temperature (Tg) of the material and the melt temperature (Tm) of the material from the material’s spectroscopic data (e.g. IR spectrum).

[0045] The step of determining one or more process parameters may comprise use of both (i) the at least one prediction regarding the material to be printed and (ii) knowledge of one or more characteristics of the part to be printed.

[0046] The method may comprise obtaining the spectroscopic data of a material currently being printed by a 3D printer (an in-line process), and that data may be used to determine the process parameters. The method may further comprise automatically measuring extrusion width of an extrudate being printed by the 3D printer, by: obtaining at least one visual image of the extrudate whilst a printing process is ongoing; performing instance segmentation on the visual image (using machine learning) to localise and segment the extrudate; performing distance measurement with reference to a calibration standard to determine the extrusion width of the identified printed extrudate; comparing the determined extrusion width to an expected extrusion width; and adjusting one or more process parameters if a difference between the determined extrusion width and the expected extrusion width exceeds a set tolerance.

[0047] According to a second aspect of the invention, there is provided a 3D printer with automatic printing process parameter determination capabilities, the 3D printer comprising: a vibrational spectrometer arranged to obtain an IR spectrum of a printing material within the 3D printer; a processing module, the processing module being arranged to: use a machine learning model to predict at least one of the identity of the material to be printed, and one or more thermophysical properties of the material to be printed, from the spectroscopic data, the machine learning model having been trained on a dataset of spectroscopic data of sample materials; and determine one or more process parameters for the printing process using the material to be printed based on the at least one prediction; and a controller, the controller being arranged to control the 3D printer in accordance with the one or more determined process parameters. The spectrometer may be an IR spectrometer. The IR spectrometer may comprise an Attenuated Total Reflection - ATR - probe. ATR is a sampling technique which is used in conjunction with infrared spectroscopy and enables samples to be examined directly in the solid or liquid state without further preparation. An FFF filament, material for melt eHD, or solution for solution eHD may therefore all be examined using the same probe. It will be appreciated that any suitable probe known in the art may be used in other embodiments.

[0048] A probe of the spectrometer may be located so as to take a spectrum of the printing material either:

[0049] (i) in a storage location of the printing material within the 3D printer; or

[0050] (ii) at a location along a travel path of the printing material between the storage location and a printing head of the 3D printer. For example, for a spooled filament to be printed, the probe may be located at the spool, or at any convenient point along the travelling filament length between the spool and the printing head.

[0051] The 3D printer of the second aspect may be used to implement the method as described with respect to the first aspect.

[0052] According to a third aspect of the invention, there is provided a method of automatically measuring extrusion width of an extrudate printed by a 3D printer, the method comprising: obtaining at least one visual image of the extrudate whilst a printing process is ongoing; performing machine learning-based instance segmentation on the visual image to localise and segment the extrudate, so identifying the printed extrudate within the visual image; and performing distance measurement with reference to a calibration standard to determine the extrusion width of the identified printed extrudate.

[0053] The combined use of localisation and segmentation is often referred to as instance segmentation (a general name for an approach combining both localisation and segmentation) - the method may therefore be said to comprise performing instance segmentation to identify the printed extrudate. Embodiments may therefore apply segmentation and / or localisation through machine learning methods to identify the printed extrudate within a visual image without human input. This segmentation may be referred to as visual segmentation, as visual data are the input for both training of the machine learning model used, and in practice. Once identified, its width can then be easily measured by reference to a pre-calibrated standard. Use of a machine learning (ML) model for instance segmentation therefore allows automatic / autonomous determination of the width of the printed extrudate.

[0054] Defects in 3D printed parts often stem from inaccuracies in one or both of extrusion height and extrusion width. Extrusion height can already be easily measured, using multiple known techniques (e.g. inductive or capacitive proximity sensors, and BLTouch sensing), but accurate measurement of extrusion width has remained elusive. Embodiments of the invention therefore provide an intelligent in situ visual monitoring and measurement system which allows extrusion width to be accurately determined in real-time, and without contacting (and so risking disruption of) the extrudate. A rapid two-stage image processing approach is used in the specific examples described to identify the extrudate to be measured within the visual image, and to separate the extrudate to be measured from the background In particular, multiple closely-spaced lines of extrudate may be present within some images, so correct localisation of the extrudate to be measured (generally the most recently -deposited portion) - and not just of any objects within the image - is needed, as well as segmentation of foreground objects from the background so as to define the bounds of what is to be measured.

[0055] Whilst some commercial 3D printers incorporate ‘sensing elements’ such as nozzle or print bed cameras, these generally rely on human monitoring and intervention in the case of print faults and failures. By contrast, embodiments of the invention provide automated defect detection, and optionally also automated pausing or termination of the printing process, or automated correction, if a defect arises. Image processing may therefore be used in combination with machine learning to provide an intelligent monitoring tool to measure the geometry of the smallest building blocks during the 3D printing process. Other approaches still in a research and development phase have some automatic error detection, but these are focused on inter-layer defects. Unlike approaches monitoring large inter-layer defects, the approach herein of looking specifically at width of each layer of extrudate as it is deposited provides much faster feedback, so facilitating smart control of the 3D printer. Moreover, systems based on detecting inter-layer defects generally employ classification methods to identify the defect, so their feedback is based on qualification of the print using discrete labels such as “good”, “normal”, or “bad”, which introduce ambiguity and subjectivity into control process. By contrast to defect-monitoring approaches, the current approach focusing on extrudate width in particular has a small tolerance for accepted error, detecting printing errors as the printer extrudes each line based on continuous measurement, eliminating the reliance on qualifying labels. This approach provides more straightforward and reliable feedback for the control process, with realtime comparison of measured and desired width values.

[0056] The method may further comprise performing a perspective correction as part of the distance measurement. The distance and perspective calibration may be integrated into a single calibration paradigm. The system used to perform the method may be pre-calibrated using a standard (a calibration standard), e.g. a marker of known dimensions. Once the camera angle of a camera arranged to provide the visual images for use in the method is fixed with respect to a printing head, the marker with known dimensions may be placed in a set position where, or near where, the extrudate to be measured will be - e.g. in front of a nozzle of the printing head, on a printing bed onto which the extrudate is to be printed in due course such that the extrudate will be in at least substantially the same plane as the marker. A transformation may then be calculated to map features (e.g. corners) of the marker as seen in the obtained visual image to a frame, the frame having the known dimensions of the marker. As the dimensions of the marker are known, angle correction and absolute dimensions can therefore be calculated for other objects placed in the same area of the printing bed. In the real-time application, the pre-determined transformation and distance calibration are used to correct the perspective of a visual image of an identified extrudate and to determine the absolute dimensions (in particular, width) of the extrudate. Application of this method may be easiest for Fused Filament Fabrication (FFF) 3D printing, however the same principles are applicable to melt- and solution- eHD. A more powerful camera / lens may be needed for use with eHD 3D printing due to the extruded fibres generally being thinner. In electrospinning modes, the visual image of the fibre may be taken immediately after detachment from the nozzle / close to a nozzle tip rather than on a printing bed, as the fibre will start to spin on approaching the bed, making it harder to track. The calibration standard used, and the camera, may therefore both be differently-positioned for electrospinning 3D printing modes. The method may therefore comprise a calibration step performed using a calibration standard of known dimensions. The calibration step may comprise, prior to obtaining the visual image of the extrudate: obtaining a visual image of the calibration standard whilst the calibration standard is located in the same plane as the extrudate to be measured will lie on being printed / extruded; and computing an affine transformation to map the image of the calibration standard to a frame with the known dimensions of the calibration standard. The same transformation and scaling can then be applied to the extrudate to be measured. The frame may provide a width value for each pixel of the transformed image, so allowing the width to be determined from the visual image. The transformation, e.g. an affine transformation, may serve to correct for perspective. In some embodiments, for example in embodiments in which a camera used to obtain the visual image directly faces a plane in which the extrudate will lie, no transformation may be needed. The calibration standard may simply be used to convert a number of pixels to an absolute width measurement in such embodiments.

[0057] The extrusion width may be determined in real-time, and optionally within 0.01 to 0.21 seconds of the measured portion of the extrudate being printed, and optionally within 0. 125 seconds of the measured portion of the extrudate being printed.

[0058] The localisation may be performed using at least one of:

[0059] (i) a deep object detection neural network architecture such as YOLO - You Only Look Once - object detection methodology (it will be appreciated that various different versions of YOLO exist with slightly differing capabilities, and that any suitable YOLO family member may be used); and

[0060] (ii) region of interest (ROI) object detection (e.g. object detection based on a constant box approach

[0061] - see e.g. the following paper for more details on a predefined ROI approach to localisation - Q. Zhang and H. Xiao, "Extracting Regions of Interest in Biomedical Images" , 2008 International Seminar on Future BioMedical Information Engineering, Wuhan, China, 2008, pp. 3-6, doi: 10.1109 / FBIE.2008.8). In embodiments using a neural network architecture, the architecture may be or comprise a convolutional neural network (CNN).

[0062] The segmentation may be performed using at least one of:

[0063] (i) a deep neural network architecture, such as UNet, ENet, or a fully convolutional neural network

[0064] - FCN;

[0065] (ii) a clustering method such as Gaussian Mixture Model - GMM - segmentation; and

[0066] (iii) colour filter segmentation.

[0067] In embodiments using a neural network architecture, the architecture may be or comprise a convolutional neural network (CNN). For example, the detection of the printed extrudate may be performed by instance segmentation methods, such as a two-stage method mixing YOLO object detection with a fully convolutional neural network - FCN. In some embodiments, the localisation may be performed by YOLO object detection and the segmentation may be performed by a fully convolutional neural network - FCN.

[0068] It will be appreciated that there are various instance segmentation approaches, any of which may be implemented for the method described herein: One-stage; Two-stage; Multi-stage; and Weakly supervised. Instance segmentation can be performed using single-stage, two-state, or multi-stage models. Here, the “stages” refer to the separated sections of the instance segmentation model - for example separate stages trained to serve for either localisation or segmentation for a two-stage model, and to one or more additional sections to extract further related information for a multi-stage model. In two- and multi-stage models, all sections of the model can be trained simultaneously or separately. The one-stage model is different in that both localisation and segmentation are integrated into a single architecture. In contrast, two and multi-stage models rely on a hierarchy of these sections, whether localisation or segmentation comes first. For example, in a two-stage implementation, the step of performing instance segmentation on the visual image to localise and segment the extrudate may comprise: performing image processing on the at least one obtained visual image to localise the extrudate to be measured; segmenting the at least one obtained visual image to distinguish objects from a background and to provide a boundary of detected objects; and combining the extrudate localisation information with the segmentation to identify the printed extrudate.

[0069] The method may comprise using the segmentation to generate a mask of the one or more distinguished objects. The mask may then be used in determining the extrudate width, by reference to the calibration standard.

[0070] The method may comprise comparing the determined extrusion width to an expected extrusion width, and taking action if a difference between the two exceeds a set tolerance. For example, the method may comprise stopping printing of the extrudate if the difference between the two exceeds a set tolerance. Alternatively or additionally, the method may comprise, in response to determining that the difference between the two exceeds a set tolerance, determining adjusted printing process parameters to correct the discrepancy. For example, the method may comprise calculating correction values for slicing parameters such as printing wall thickness. A processing module may calculate one or more correction parameters that can be used in the slicer to generate optimised G-code (i.e. printing process controls interpretable by an electronic controller of a 3D printer) with respect to the current conditions of the 3D printer (the “slicer” being the general name for the 3D printing software that acts as a link between the digital model of a part to be printed and the actual model being printed). In some embodiments, the method may comprise automatically adjusting the printing parameters to correct for a width discrepancy if the difference between the determined extrusion width and the expected extrusion width is below a pre-set threshold, and stopping the printing process is the difference is above the same, or a different, pre-set threshold. Different actions may therefore be taken depending on the extent of the discrepancy.

[0071] At least five, and optionally at least ten, visual images per second may be obtained and processed according to the method of this third aspect. In various embodiments, the image processing capabilities may vary between 100 frames / second and 5 frames / second. In various embodiments, the method may be arranged to provide measurements for a 10 mm length of extrudate from each captured frame, and the speed of measurement varies may be between 50 mm / s and 1000 mm / s depending on specifics of the selected approach. The method of the third aspect may be used in conjunction with the method of the first aspect.

[0072] According to a fourth aspect of the invention, there is provided a 3D printer with integrated extrudate width measuring capabilities, the 3D printer comprising: a printing head comprising a nozzle through which the extrudate is printed; a camera mounted on the printing head (optionally via a camera mount, which may be adjustable, mounted on the printing head) and arranged to provide visual images of the printed extrudate; and a processing module arranged to: obtain at least one visual image from the camera; perform machine learning-based instance segmentation on the visual image to localise and segment the extrudate, so identifying the printed extrudate within the visual image; and perform distance measurement with reference to a calibration standard to determine the extrusion width of the identified printed extrudate.

[0073] The processing module (which may also be referred to as a processing unit) may be arranged to compare the determined extrusion width to an expected extrusion width. The 3D printer may further comprise a controller arranged to receive a signal from the processing module and to take action if a difference between the two widths exceeds a set tolerance - for example to stop printing of the extrudate if a difference between the two widths exceeds a set tolerance, or to make corrections to one or more process parameters so as to correct the extrudate width without stopping printing

[0074] The 3D printer may comprise a light arranged to illuminate the printed extrudate. The printer may further comprise a light mount arranged to mount the light on the printing head, optionally adjustably.

[0075] The 3D printer may further comprise a shroud mounted on the printing head, the shroud having the camera and optionally a light mounted thereon. The shroud may further comprise the camera mount and / or light mount, where present. The one or more mounts may be integrally formed with the shroud, or separately made and mounted thereon In embodiments with a shroud, the shroud may be rigidly mounted on the printing head such that the shroud has a set position with respect to the nozzle. The shroud may be integrally formed with the printing head in some embodiments, however it will be appreciated that a modular design may be favoured in some embodiments for both adjustability and cost-saving reasons.

[0076] The camera mount and / or the light mount, where present, may be adjustable such that a desired image angle / location is selected, and well-illuminated where a light is used. This adjustment may be performed prior to printing, and prior to a calibration step. In some embodiments, the mount(s) may not be adjusted in use, such that light and camera angles and positions are constant throughout use. In other embodiments, the mount(s) may be adjustable in use, optionally between two or more pre-set positions for which calibration has already been performed to allow distance measurement. A processing module may select between different sets of known calibration information (e.g. affine transformation and pixel width value information sets) depending on the camera position.

[0077] In a two-stage or multi-stage process, the processing module may be arranged to, in the step of performing instance segmentation: perform image processing on the at least one obtained visual image to localise the extrudate to be measured; segment the at least one obtained visual image to distinguish objects from a background and to provide a boundary of detected objects; and combine the extrudate localisation information with the segmentation to identify the printed extrudate. The localisation and segmentation may be performed in either order, or simultaneously.

[0078] The 3D printer of the fourth aspect may be used to implement the method as described with respect to the third aspect. The 3D printer of the fourth aspect may further comprise vibrational spectrometer, e.g. an IR spectrometer, and may be made to function as the 3D printer of the second aspect.

[0079] According to a fifth aspect, there is provided a multi-modal 3D printer arranged to be switchable between a fused filament fabrication - FFF- 3D printing mode and at least one electrohydrodynamic - eHD - 3D printing mode, the printer comprising: a printer frame arranged to support other components of the 3D printer; a printing head comprising at least one nozzle with electrical grounding; an insulating printing bed apparatus comprising: a conductive plate arranged to be electrified when the printer is operated in the eHD 3D printing mode, the conductive plate being arranged to provide a printing bed surface in use; an electrically-insulated layer; an electrical insulation frame surrounding the edges of the conductive plate and arranged to isolate the conductive plate from the printer frame and to mount the conductive plate on and parallel to the electrically -insulated layer, between the printing bed and the printing head, a thickness of the electrical insulation frame spacing the conductive plate from the electrically-insulated layer; and a spacer mounted on the opposite side of the electrically-insulated layer from the conductive plate, and comprising at least one bed-levelling adjustor; and a motion carriage arranged to move the insulating printing bed apparatus, wherein the spacer is arranged to distance the conductive plate from the motion carriage, and the electrical insulation frame is arranged to isolate the conductive plate from the motion carriage.

[0080] As the conductive plate is electrifiable, it can act as both standard printing bed surface for FFF printing, and as the collector for eHD printing. The term “printing bed” is therefore used for ease to cover both. Embodiments of the invention therefore provide a 3D printer capable of continuous transition between 3D- printing modes - for example between a Fused Filament Fabrication (FFF) 3D printing mode and one or more electrohydrodynamic (eHD) 3D printing modes. The ability to switch between modes smoothly within the same printer may allow a user to print structures containing wide ranges of feature size (e.g. centimetre to micrometre), shape, and structure, by automatically changing of the printing mode within a single printing process. The integrated design of such an FFF-eHD multi-modal 3D printer may therefore increase freedom of the additive manufacturing design in term of material, structure, and process, by smoothly changing the printing modes without interrupting the printing process.

[0081] The insulating printing bed apparatus may further comprise an electrically insulated motion carriage plate, the motion carriage plate being arranged between the spacer and the motion carriage and used to connect to the motion carriage of the 3D printer.

[0082] The spacer may be made from an electrically insulating material. The spacer may be arranged to distance the conductive plate from the motion carriage by a distance of at least 5 mm, and optionally by around 45 mm. It will be appreciated that a greater spacing may reduce the risk of current leakage. Spacer height may be selected based on dielectric strength of the spacer material and the maximum expected applied voltage, among other factors.

[0083] The 3D printer may be arranged to be switchable between at least two different electrohydrodynamic - eHD - printing modes. For example, the 3D printer may be arranged to operate in two or more of:

[0084] (i) a melt electrospinning mode;

[0085] (ii) a solution electrospinning mode; and

[0086] (iii) an electro-spraying mode.

[0087] Different modes may require different nozzles - for example, melt electrospinning may be performable using an FFF material extrusion nozzle (provided the nozzle is electrically grounded), whereas a different nozzle specifically designed for spraying a more liquid material to be printed may be more suitable for a solution electrospinning mode or electro-spraying mode - such a nozzle may be referred to as a spraying nozzle or spinneret. A spraying nozzle / spinneret, which may also be referred to as a solution nozzle, may comprise a needle, although other geometries and designs are also available. The printing head may therefore comprise at least two grounded nozzles; the at least two grounded nozzles comprising a spraying nozzle and a melt nozzle. For example, the 3D printer may be arranged to be switchable between at least two electrohydrodynamic - eHD - printing modes; a melt electrospinning mode and one or more of a solution electrospinning mode and an electro-spraying mode, and the printing head may comprise at least two grounded nozzles: the at least two grounded nozzles comprising a spraying nozzle and a melt nozzle. More generally, any suitable eHD techniques such as droplet on demand, or flow focusing, may be implemented instead of, or as well as, other eHD techniques. In embodiments with a solution nozzle which is or comprises a spinneret (e.g. a metal capillary or needle), the 3D printer may further comprise a spinneret grounding apparatus The spinneret grounding apparatus may comprise two conductive contacts biased towards each other, the biasing direction being perpendicular to the length of the spinneret. The two conductive contacts may be arranged on opposing sides of the spinneret from each other such that they are biased to grip the spinneret therebetween, and each conductive contact is electrically grounded. In some embodiments, both conductive contacts may be movable, for example both being mounted on springs and movable along a single axis under the influence of the respective spring (or other biasing member). In other embodiments, one may be static / rigidly held in place, and the other may be biased towards the static contact. The movable, biased, contact(s) may allow the same spinneret grounding apparatus to be used with a variety of differently-sized needles, capillaries, or other spinneret designs. The grounding apparatus may comprise a pair of singleacting spring return cylinders on opposite sides of the spinneret, each cylinder having a conductive contact biased towards the spinneret. Alternatively, the grounding apparatus may comprise a pair of contacts each connected to a supporting structure by at least one biased connector arm, the connector arms optionally being rotatably mounted on both the supporting structure and the contact, so allowing a contact face of the contact to remain parallel to the spinneret whilst moving perpendicularly to the spinneret to accommodate different spinneret widths.

[0088] In embodiments with multiple nozzles, each nozzle may have a different effective surface area. An electrically insulating wall may be provided to avoid charge transfer between the two nozzles. The electrically insulating wall may prevent charge transfer between the two nozzles even at high applied voltages (> 15 kV). In embodiments where one or more nozzles has a needle or other spinneret, the electrically insulating wall may be located around the spinneret, and optionally may be provided by a portion of a spinneret grounding apparatus support structure. More generally, the electrically insulating wall, where present, may surround at least part of a particular nozzle, or form a barrier between nozzles.

[0089] The 3D printer may be arranged to be switchable between an FFF 3D printing mode and a melt electrospinning 3D printing mode, and to use the same nozzle for both FFF 3D printing and melt electrospinning. If the rheology of the material to be printed allows (e.g. when using a polymer with low viscosity), electrospraying may be performed with the same nozzle.

[0090] The conductive plate may be a conductive metal plate, e.g. aluminium, steel, or any suitable pure metal or alloy. The conductive plate may be between 2 mm and 4 mm thick, and optionally may be 3 mm thick.

[0091] The electrically-insulated layer may be formed from a standard FFF printing bed which has been electrically insulated with electrically-insulating plastic. For example, the insulation may be provided in the form of two insulating film layers, one on each face of the generally planar printing bed. Optionally, each layer may be a mica plastic film. The pre-existing printing bed may therefore form at least part of the electrically-insulated layer as described above - optionally, the pre-existing printing bed together with one or more layers of electrically-insulating plastic applied thereto may form the electrically-insulated layer as described above.

[0092] The electrical insulation frame may be made of a compliant elastic material, and may be arranged to reversibly deform so as to be securely fitted to the electrically-insulated layer, which may or may not be formed from a pre-existing printing bed of a 3D printer. The use of a deformable, elastically-compliant, material to grip the electrically-insulated layer may serve to avoid a need for screws, or other fasteners - it will be appreciated that metal screws in particular could interfere with the electrical conduction / insulation properties desired, and in particular could lead to increased current leakage. The electrical insulation frame may also be arranged to reversibly deform so as to be securely fitted to the conductive plate, again avoiding the need for fasteners.

[0093] The spacer may be mounted on an axis motion plate of the motion carriage. The electrically-insulated layer, and the printing bed apparatus more generally, may be rectangular (e g. square) in shape, and the spacer may comprise four sections, one located in each corner region of the rectangular printing bed apparatus / layer. The at least one bed-levelling adjustor may comprise a bed-levelling spring and a wheel arranged to adjust the spring. Optionally, four may be provided - one in each corner region of the bed.

[0094] The 3D printer may further comprise a current sensor arranged to sense a current flowing within printing control circuitry of the 3D printer, and a processing module arranged to receive data from the current sensor. The processing module may be arranged to infer in which printing mode the printer is currently operating from the current, and to stop the printing process if there is a mis-match between the inferred printing mode and an intended printing mode, for example as described above with respect to the first aspect.

[0095] The 3D printer may comprise a vibrational spectrometer, for example an IR spectrometer, and may be arranged to perform the method of the first aspect, or to act as the 3D printer of the second aspect. The 3D printer may comprise a camera arranged to obtain a visual image of an extrudate, and may be arranged to perform the method of the third aspect, or to act as the 3D printer of the fourth aspect.

[0096] According to a sixth aspect of the invention, there is provided a method of retrofitting a printing bed of an FFF 3D printer to allow usage of the printing bed in at least one electrohydrodynamic 3D printing mode, the method comprising: coating each face of the printing bed with an insulating layer to form an electrically-insulated layer; mounting a conductive plate on and parallel to the electrically-insulated layer using an electrical insulation frame which surrounds the edges of the conductive plate, a thickness of the electrical insulation frame spacing the conductive plate from the electrically-insulated layer; and mounting a spacer on the opposite side of the electrically-insulated layer from the conductive plate, the spacer comprising at least one bed-levelling adjustor; and wherein the spacer is arranged to distance the conductive plate from a motion carriage of the FFF 3D printer, and the electrical insulation frame is arranged to isolate the conductive plate from the motion carriage.

[0097] The step of coating each face of the original printing bed with an insulating layer may comprise applying a polymeric film (e g. mica plastic film) to each face of the (generally planar) printing bed. The conductive plate may form the upper layer of the assembly, above the original printing bed (in a standard 3D printing orientation), such that the material is printed onto the conductive plate - an upper surface of the conductive plate may therefore be the surface onto which the extrudate is deposited.

[0098] The electrical insulation frame may be made of a reversibly deformable material, and the step of mounting of the conductive plate on the electrically -insulated printing bed may comprises deforming the electrical insulation frame so as to fit the conductive plate and the electrically-insulated layer therewithin, and then releasing the electrical insulation frame. The electrical insulation frame may be shaped so as to provide separate, spaced, positions for the conductive plate and the electrically-insulated layer. The spacer, and optionally also the electrical insulation frame, may themselves be made by 3D printing, so allowing for precise customisation and tailoring of the shape and size to specific printing beds and use-cases. This may also reduce fabrication costs. The method of the sixth aspect may create an insulating printing bed apparatus as described in the fifth aspect.

[0099] The various aspects of the invention may be used in combination in some embodiments, yielding a highly efficient, responsive / intelligent, multi-resolution, multi-mode, multi-dimensional printing technology. Such embodiments may use both in-line characterisation for printing parameter selection and intelligent in situ visual monitoring and measurement for correction whilst printing. The technology can be used for polymer- based manufacturing (including composites), biological solutions, and pharmaceuticals. The variety of structures and high resolution provided by this technology means that such a printing system may be used to manufacture a wide variety of products, e.g. for applications including energy, healthcare (tissue engineering), textiles (fashion), construction, pharmaceuticals, biosensing, microrobots, food (encapsulation and printing foods), packaging, 2D materials (face masks, wound dressings, PPE, membranes), catalysis, wastewater, electrical circuitry etc. By contrast, prior art multi-modal printers are not capable of smoothly incorporating different printing modes continuously within a single print work. Since these printers use specific components for each printing mode, such as a pressure-based feeding system and a needle-syringe extrusion printing head, they require a complete reconfiguration of the entire setup to switch between printing modes. Moreover, these devices cannot perform solution electrospinning (SES) and electrospraying as well as melt electrospinning (MES) and Fused Filament Fabrication (FFF, i.e. “traditional”) 3D printing. Embodiments of the invention provide a 3D printer that is capable of mode switching between any and all of these modes during the fabrication process. Moreover, process parameters for each mode can be set independently based on results of in-line characterisation of the material to be printed. The multi-modal printer is able to deposit structures and materials with a wide range of geometries, morphologies, and resolutions, by switching between printing modes on demand.

[0100] Features described with respect to one aspect of the invention may be applied to any other aspect of the invention, mutatis mutandis.

[0101] There now follows by way of example only a detailed description of embodiments of the present invention with reference to the accompanying drawings in which:

[0102] Figure 1 shows a schematic diagram of a multi-modal 3D printer for use with any aspect of the invention; Figure 2 shows a schematic diagram of a 3D printer including a differently -positioned IR spectrometer;

[0103] Figure 3 is a plot showing predicted glass transition temperature of polymers using a random forest regressor, alongside the outcome of classification for individual points using random forest classification; Figure A is a bar chart showing the performance of material identity classification using the IR spectra;

[0104] Figure 4B is a bar chart showing the error associated with predictions of Tgfrom regression models;

[0105] Figure 5 illustrates a method of generating process parameters for a 3D printing process;

[0106] Figure 6 provides a close-up cross-sectional view of a nozzle printing a material onto a printing bed;

[0107] Figure 7 provides multiple photographs of a printing head arranged to measure extrudate width;

[0108] Figure 8 shows a system for in-line measurement of a 3D-printed extrudate;

[0109] Figure 9 provides an overview of a computational approach for detection and measurement of an extruded filament, comprising localisation and segmentation methods;

[0110] Figure 10 illustrates use of the width measurement algorithm of Figure 9 for three printing scenes (visual images) of increasing complexity; Figure 11 shows plots of the SMAPE of width measurement using semantic segmentation models with (a) predefined ROI (constant box) detection, and (b) YOLO methodology;

[0111] Figure 12 shows a plot demonstrating the sensitivity of various combinations of methodologies to the airgap size between printed roads;

[0112] Figure 13 shows a plot of speed vs. accuracy for two-stage instance segmentation methods implemented on an NVIDIA Tesla T4 GPU;

[0113] Figure 14 illustrates a calibration marker for use in determining extrudate width;

[0114] Figure 15 illustrates a method of measuring extrudate width in-line in a 3D printing process;

[0115] Figure 16A shows a shroud used to mount components used for the measuring method on a printing head; Figure 16B shows the shroud of Figure 16A on the printing head, with other components mounted thereto; Figure 16C shows the apparatus of Figure 16B from a different angle;

[0116] Figure 17 shows the components of the camera mounts shown in Figures 16B and 16C;

[0117] Figure 18 shows the components of the light mounts shown in Figures 16B and 16C;

[0118] Figure 19 shows the printing head with shroud in context;

[0119] Figure 20 shows an insulating printing bed apparatus that can provide a suitable printing bed for both FFF and eHD 3D printing;

[0120] Figure 21 shows the insulating printing bed apparatus of Figure 20 with two components removed;

[0121] Figure 22 shows a cross-sectional view of the insulating printing bed apparatus of Figure 20;

[0122] Figure 23 shows close-up views of the spacer of the insulating printing bed apparatus of Figure 20;

[0123] Figure 24 shows photographs of an insulating printing bed apparatus formed on a commercial 3D printer;

[0124] Figure 25 illustrates a method of retrofitting an insulating printing bed apparatus to a 3D printer;

[0125] Figure 26 illustrates a printing head with a melt nozzle and a solution nozzle, and a spinneret grounding apparatus;

[0126] Figure 27 illustrates the spinneret grounding apparatus of Figure 26;

[0127] Figure 28 provides three photographs of a printing head in accordance with Figure 26;

[0128] Figure 29 provides two perspective views of the printing head of Figure 26;

[0129] Figure 30 provides close-up views of the support structure of the grounding apparatus of Figure 26;

[0130] Figure 31 provides close-up views of other components of the spinneret grounding apparatus of Figure 26;

[0131] Figures 32 and 33 illustrates a printing head with an alternative spinneret grounding apparatus;

[0132] Figure 34 illustrates a support structure of the spinneret grounding apparatus of Figure 32;

[0133] Figure 35 illustrates other components of the spinneret grounding apparatus of Figure 32;

[0134] Figures 36A and 36B illustrate the spinneret grounding apparatus of Figure 32 in situ on the printing head, showing the springs of the spinneret grounding apparatus, in perspective and partially cut-away views, respectively; and

[0135] Figure 37 illustrates a cap holder of the solution nozzle shown in Figures 32 and 33.

[0136] Figure 1 shows a 3D printer 100, and in particular a multi-modal 3D printer 100 which is capable of printing in three different printing modes - FFF, solution eHD, and melt eHD 3D printing. The 3D printer 100 comprises a printing bed 200, also referred to as a print bed 200. The printing bed 200 provides a flat surface on which the 3D printer builds an object. In eHD 3D printing, the printing bed 200 is often referred to as a collector 200 - the terms “printing bed” and “collector” may therefore be used interchangeably herein as the same structure is used as both the “collector” 200 for eHD printing and the “printing bed” 200 for FFF printing. The 3D printer 100 comprises a printing head 150, also referred to as a print head 150. Material to be printed is arranged to pass through the printing head 150 so as to be deposited on the printing bed 200 and build up the printed object. The 3D printer 100 is arranged to provide controlled relative movement between the printing bed 200 and the printing head 150. Either or both of the printing bed 200 and the printing head 150 may be arranged to move. In the embodiment being described, the printer 100 further comprises a motion carriage 120, and the printing bed 200 is mounted on the motion carriage 120. The motion carriage 120 is arranged to move the printing bed 200 along two axes / within a plane (generally a horizontal plane). In the embodiment being described, the printing head 150 moves vertically only, along a frame 101, and movement of the bed 200 provides all horizontal movement. In alternative embodiments, all movement may be provided by movement of the printing head 150, all movement may be provided by movement of the printing bed 150, or both the printing head 150 and the printing bed 200 may provide movement on one or more axes, and optionally on all axes.

[0137] The printing head 150 comprises at least one nozzle 160, 170. In the embodiment being described, the printing head comprises two nozzles - a first nozzle 160 designed for printing extruded filaments in FFF 3D printing and a melt material in melt eHD, and a second nozzle 170 designed for spray printing, e.g. for solution eHD. More or fewer nozzles may be provided in other embodiments. In embodiments in which one or more nozzles 160, 170 are intended to be used for eHD printing, those nozzles are electrically grounded as a current is applied to the printing material. In the embodiment shown, both nozzles are electrically grounded.

[0138] In some embodiments, especially embodiments in which eHD is to be performed, a chamber (which may be referred to as an eHD chamber) is provided, enclosing the entire printing setup 100. Factors such as environmental temperature and humidity can strongly affect electrospinning processes and, in some cases (generally for specific filaments), can also affect FFF printing quality. Humidity and temperature can be controlled inside the chamber by using a humidifier and / or dehumidifier and a heater. One or more fans may be provided for cooling. The chamber can protect users from pollutants in the air caused by solvent evaporation, and also protect the printing process from external pollutants, or indeed accidental damage.

[0139] The 3D printer 100 further comprises a store 102, 104, 106 for each printing material. For example, a spool 102 of a filament for FFF 3D printing, a spool 104 of a material for melt eHD, and a reservoir or tank 106 of a solution for solution eHD are all provided in the embodiment pictured in Figure 1. The same material may be used for both FFF and melt eHD in some embodiments. More or fewer materials may be stored in other embodiments, and / or different selections of materials may be stored - for example, multiple spools 102 of different filaments for FFF 3D printing and / or melt eHD may be provided. In other embodiments, a powder or granule store 102, 104 may be provided for a polymeric material for use in FFF or melt eHD 3D printing instead of a filament spool, and the feed mechanism 161 to the nozzle 160 may be adjusted as appropriate. The 3D printer 100 is arranged to feed whichever printing material is to be printed from its store 102, 104, 106 to the printing head 150, and more specifically to the relevant nozzle 160, 170 for that printing material. The printing material 10a, 10b (i.e. the material intended to be used in a current printing process to make at least a portion of the object to be printed) may be heated or cooled as it travels along its path to the respective nozzle outlet 160a, 170a. The heating and / or cooling may be applied between the store 102, 104, 106 and the printing head 150, or within the printing head (e.g. in a nozzle 160, 170).

[0140] The 3D printer 100 further comprises a processing module 110 and a controller 112. The processing module 110 and controller 112 may both be implemented in software, with associated circuitry. In various embodiments, either or both of the processing module 110 and the controller 112 may be provided by a computer to which the 3D printer 100 is connected. In the embodiment being described, the processing module 110, controller 112 and a user interface 108 are all provided by a laptop computer connected to the 3D printer 100. A separate user interface unit 108 may be provided in other embodiments.

[0141] The processing module 110 is arranged to receive data regarding the printing process, and optionally also data regarding the part to be printed and its desired properties, and to make decisions regarding control of the printing process. The processing module 110 may comprise a memory with stored data and / or algorithms for use in the decision-making processes, and in particular may comprise one or more machine learning models in some embodiments, as described in more detail below. The decisions may be fed to the controller 112 to be put into effect. Multiple data lines 110a are provided to transmit information from other parts of the system 100 to the processing module 110. A monitoring system 111, optionally including one or more cameras, one or more current sensors, and / or one or more temperature sensors, is also included in the 3D printer 100 and arranged to supply data to the processing module 110 in the embodiment pictured. The controller 112 is arranged to adjust parameters of an ongoing printing process - e.g. feed rate of the printing material, by how much the printing material is heated, or a current or voltage applied to the printing material - and to control the relative position of the printing head 150 and printing bed 200. Multiple command lines 112a are provided to transmit commands from the controller 112 to other parts of the system 100. The data and command lines 110a, 112a may be wired or wireless, and any suitable communication protocol may be used. For example, for the melt material 10a, the controller 112 may control the rate of supply / feed rate of the printing material 10a to the nozzle 160 by means of a plastic feeder 161 of adjustable speed, and may control the temperature of the printing material 10a reaching the nozzle 160 by means of a controllable heater 162. The heater 162 is located within the printing head 150 in the embodiment shown, so as to adjust the heat to a desired level close to the point of use. In 3D printers 100, the entire extrusion set-up 162, 160, 160a on the print head 150, including the nozzle 160, heating block 162 or other heater, any thermocouple(s), and any heat sink is generally referred to as the “hot-end”. A general temperature of the hot-end, and / or a specific temperature at the heating block 162 or in the nozzle 160 may be provided to the processing module 110. Similarly for the solution 10b, solution extrusion equipment 173, e.g. a pneumatic pump or screw plunger, is provided to control the feed rate of the material 10b, and may be controlled by the controller 112 (although no control line 112a is shown for this in Figure 1 for clarity).

[0142] The 3D printer 100 also comprises a power supply 116, the power supply being a DC voltage power supply in the embodiment being described, and a power supply control unit 117. A two-way data transfer and command transfer link is provided between the power supply control unit 117 and the processing module 110 and controller 112 of the 3D printer.

[0143] In the embodiments shown in Figures 1 and 2, the 3D printer 100 further comprises an IR spectrometer 114 with an Attenuated Total Reflection - ATR - probe. It will be appreciated that different IR spectrometers 114, and / or vibrational spectrometers of other kinds - e g. Raman spectrometers, may be used in different embodiments, and that IR probes and spectra are used as the example herein for ease of explanation only - this is not limiting on the scope of the invention. The IR probe 114 is arranged to record an infra-red spectrum of the printing material. IR radiation generated at the probe 114 is directed at the printing material (e.g. a polymer or filament 10a, or an eHD solution 10b); some radiation is absorbed and some passes through (is transmitted). The resulting signal at a detector of the IR spectrometer is a spectrum representing a molecular ‘fingerprint’ of the material.

[0144] In the embodiment shown in Figure 1, the IR probe 114 is located between the material store 102, 104, and the printing head 150, and arranged to record a spectrum as the printing material 10a moves towards the printing head 150. In the embodiment shown in Figure 2, the IR probe 114 is located on or adjacent to the printing head 150, and arranged to record a spectrum of the solution-state printing material 10b in the reservoir 106. The probe 114 may be located within the reservoir 106 in some embodiments. In other embodiments, the IR probe 114 may be differently -located, for example so as to record a spectrum of material still in the store 102, 104, 106, or of material within the printing head 150, or of material as it is ejected from the nozzle 160, 170, or even after it has been printed. IR probe location may therefore be chosen as appropriate for a given embodiment - for example in a storage location 102, 104, 106, of the printing material within the 3D printer 100; or at any suitable location along a travel path of the printing material between the storage location 102, 104, 106 and a printing head 150 of the 3D printer 100, and multiple probes 114 in different locations may be provided in some embodiments. In addition, in some embodiments, the IR probe 114 may be movable in use - for example between a plurality of positions each corresponding to one of the multiple printing materials available to the 3D printer 100. An IR probe position may therefore be selected based on for which printing material an IR spectrum is wanted.

[0145] A mathematical correlation was developed, using machine learning, between the IR spectra of various printing materials (generally polymers) and their physicochemical properties (e.g. glass transition temperature, Tg, and melting temperature, Tm). It was demonstrated that IR spectra of an unknown polymer can be used to predict its melting temperature and / or its glass transition temperature. Knowledge of one or both of these temperatures is key to determining a suitable temperature for the nozzle 160, 170 to be used for printing the material in question / to what temperature the material should be heated for printing. Fourier Transform Infrared (FTIR) spectra were used in the specific embodiment being described, although it will be appreciated that different approaches to using IR data may be used in other embodiments - for example, near infrared (NIR) spectra may be used in some embodiments. The FTIR spectra were used as descriptors for machine learning techniques, to identify the input material and / or predict its thermophysical properties.

[0146] Machine learning techniques were applied to the FTIR spectra, in conjunction with other material data for predicting thermophysical properties in the embodiment being described - more specifically, transmission wavelengths were extracted and normalised, and these values were combined with a value for the material’s density and used as an input for regression. In other embodiments, density may not be used, and / or another material property may be provided instead of, or as well as, density, for use alongside the IR spectra. The IR spectra may be used alone in other embodiments.

[0147] The resultant algorithm’s ability to either identify the polymer type or predict the glass transition temperature (Tg) respectively from this data alone (IR spectra and optionally density) was analysed. A dataset was built from IR spectra taken from the online Wiley database, spectrabase.com; 864 individual spectra were extracted, representing 131 distinct polymers. Where only a single spectrum was available for a particular polymer, that spectrum was withheld from the training data set. Of these 131 polymers, 30 only had one associated spectrum, and these 30 spectra / polymers were therefore “unknown” to the model when included in the test set. The testing phase therefore assessed (i) how well the model could recognise “known” polymers (the 101 polymers for which a spectrum was included in both the training data and the test data) and (ii) how well the model could extrapolate its predictions for “unknown” polymers (the 30 polymers for which a spectrum was included in the test data, but not in the training data). In the tests performed for this study, all data in fingerprint region of the spectra, between 1800 cm'1and 700 cm'1was grouped into bins (binned); the average values between 1800-1250 cm'1and 1250-700 cm'1were used as descriptors alongside the material density in random forest classification, to identify the polymer type, and in random forest regression, to predict the Tgof the polymer. Two different machine learning (ML) models were tested - one for classification of a set of known materials and one for predictions of thermophysical properties for both known materials and “unknown” / “unseen” materials The ML model for classification was used to identify a material from a set of known materials, and was only tested on materials from within that known set. By contrast, regression models (to predict thermophysical properties directly) were used for both known and unknown polymers. The former (classification model) was found to correctly identify the “known” plastics from their IR spectra alone, with a high degree of accuracy. Thermophysical properties for the known materials can then be retrieved from a data store. To make successful predictions of thermophysical properties for both known and unknown materials without classification, use of an additional physical attribute of the material - e g. density - was found to be helpful. Such physical attributes can also be measured or otherwise determined in-line in the 3D printing process. The use of this additional physical attribute in conjunction with the IR spectrum was found beneficial for predictions of thermophysical properties (by regression), for both known and unknown materials. The IR spectra directly correlate to the elemental / chemical composition and bonding which is present in the macromolecule. For the IR regression to predict thermophysical properties, the IR spectra were therefore paired with another descriptor to represent the packing / sterics of the polymer (in this case, density, but other attributes such as viscosity / rheometry could be used instead or as well as density). The results of this study are shown in Figures 3, 4A, and 4B; known polymers were correctly predicted 91.66% of the time by the classification model, and predictions of glass transition temperature agreed well with synthetic values, with R2> 0.9, for both known and unknown polymers. This technology can be integrated into the 3D printer 100, to intelligently select process parameters for a given filament or solution without human intervention. Figure 3 shows a plot of the predicted glass transition temperature (Tg) of polymers using a random forest regressor against the experimentally-measured glass transition temperature, showing strong matching. Figure 4A is a bar chart showing the performance of classification / predictions of material identity, showing correct identification 92% of the time, and Figure 4B is a bar chart showing the error associated with predictions of Tgfrom regression models, illustrating that this is generally very close to zero. It will be appreciated that the reported performance is for a prototype system trained on a relatively small dataset, and that accuracy and reliability will be further refined.

[0148] Further testing of the method described herein was then performed, with a wider range of polymers and of machine learning model implementations. Raw infrared signals (860 samples, for 131 polymers - taking Tgand Tmfrom PolymerGenome and PolymerDatabase, with corresponding spectroscopy data obtained from SpectraBase) were used without preprocessing, demonstrating that the model works with raw, unfiltered data whilst still achieving strong predictive power. It will be appreciated that some preprocessing (such as baseline correction, scatter correction, and / or noise removal) may be preferred for poorer-quality data. Melting temperature, Tm, of the polymer materials included in this study spans a broad range, from under 300K to around 800 K.

[0149] The entire IR spectroscopy dataset was provided to a machine learning (ML) model for predicting the glass transition temperature (Tg) of polymers, rather than manually isolating the specific "fingerprints" for each polymer before providing the data to the ML model, and the ML model was shown to operate well on this “raw” data, so avoiding the need for additional preparatory work. Not all IR signals in the dataset have the same number of data points; thus, the wavelength data were standardised for use by the ML model. Linear interpolation was therefore applied based on maximum and minimum wavelengths to standardise the wavelength data across all polymers, followed by nearest-neighbour interpolation for estimating intensity values at common wavelengths. In the experimental tests being described, the IR signals cover a broad wavelength range from 400 to 4000 cm-1. No explicit start or end points were pre-set for individual polymers, as the entire spectrum is used for prediction. The dataset was randomly split with 80% of it being used as a training set and the remaining 20% as a testing set. The regression performance of the models was given using the metrics root mean squared error (RMSE), mean absolute error (MAE), and R2. The analysis of Tgand Tmpredictions was conducted separately, beginning with Tg.

[0150] In the process of predicting polymer temperature characteristics (Tgand Tm) from IR signals, nine machine learning methods were employed: Artificial Neural Network (ANN), Convolutional Neural Network (CNN), Support Vector Regression (SVR), Kernel Ridge Regression (KRR), Random Forest (RF), Gradient Boosting Machine (GBM), XGBOOST, CATBOOST, and K-Nearest Neighbour (K-NN) Regression. Testing a variety of models allows for a comprehensive comparison to determine the most suitable approach for accurate prediction. A diverse set of ML models (including in particular SVR, CNN, DNN, RF, and GBM) were therefore tested, with Support Vector Regression (SVR) consistently providing the most accurate predictions for Tg, achieving a relatively low prediction error of ±18.54 K (in the output dataset, Tgvalues ranged from 168K to 600 K.). Convolutional neural network (CNN) and deep neural network (DNN) models also demonstrated strong performance, with R2values as high as 0.89, although they exhibited slightly larger error margins than SVR.

[0151] Superior predictive performance was attained using the raw IR dataset for Tgprediction, as compared to using Principal Component Analysis (PCA)-applied data, demonstrating the value of retaining the full feature set from IR signals. The same approach was therefore used for Tmprediction.

[0152] Residual analysis revealed that the majority of predictions fell within the accurate range (±20K) for both Tgand Tmpredictions, especially for SVR and CNN models, underscoring their robustness. IR spectroscopy data for the polymer samples was therefore shown to provide strong predictive power for Tgprediction, particularly when raw spectral features were used. The SVR model’s performance significantly deteriorated when PCA was applied, suggesting that the original IR signal data contains important features for accurate Tgprediction. CNN models were found to be less sensitive to PCA reduction, indicating that certain ML models can adapt to feature-reduced datasets; raw data still yielded the best results, however.

[0153] Whilst IR data were used in these specific examples described, it will be appreciated that other spectroscopic data may be used in other implementations.

[0154] With knowledge of the polymer type and / or thermophysical properties such as melt temperature (Tm), a desired nozzle temperature can therefore be determined - an understanding of both the chemical and physical composition of the material influences the determination. More generally, multiple process parameters may be determined based on knowledge of the chemical and / or physical composition of the material to be printed, and optionally also using input information of the desired part to be printed.

[0155] A further machine learning model is therefore used for printing process parameter determination. This parameter determination ML model is trained on a training set of material identities and / or thermophysical properties and suitable process parameters, the training set being generated from prior (optionally manual) experimental characterisation to correlate polymer / polymer properties to process parameters. The ML model is trained on this dataset to predict the process parameters of other materials.

[0156] Figure 5 provides an overview of the method 500 of generating process parameters for a 3D printing process. The method 500 comprises obtaining 502 an Infra-Red - IR - spectrum (or any other vibrational spectrum, in other embodiments) of a material to be printed. This spectrum may be captured by an IR probe 114, and provided to a processing module 110 for processing. The method 500 may further comprise obtaining additional information about the material to be printed (in particular, data relating to its packing / sterics, e.g. its density), and providing that to the processing module 110 as well as the IR spectrum. The method 500 comprises providing 504 the IR spectrum of the material to be printed, and optionally also any data relating to the material’s packing / sterics, to a machine learning model. The ML model may be stored in the processing module 110, or in a memory associated therewith. The machine learning model has been trained on a dataset of IR spectra (and optionally also of data relating to the material’s packing / sterics) of sample materials to predict at least one of (i) the material / identity of a sample, and (ii) one or more thermophysical properties of a sample, from an IR spectrum spectra (and optionally also of data relating to the material’s packing / sterics) of that sample. The method 500 comprises predicting 506 at least one of the identity of the material to be printed, and one or more thermophysical properties of the material to be printed using the machine learning model, for example by using random forest classification to identify the material, and / or random forest regression to predict the one or more thermophysical properties (e.g. Tg). In various embodiments, the method 500 comprises predicting 506 both the identity of the material to be printed, and one or more thermophysical properties of the material to be printed, using the machine learning model. The method 500 comprises determining 508 one or more process parameters for the printing process using the material to be printed, based on the at least one prediction provided by the (first) machine learning model. The determination 508 may be performed using a second machine learning model trained on a dataset of printing material identities and / or thermophysical properties, and suitable printing process parameters associated with those identities and / or thermophysical properties.

[0157] In various embodiments, the method 500 further compromises taking action based on the one or more determined process parameters - for example, the processing module 110 may provide the determined parameters to the controller 112 of the 3D printer, and the controller 112 may adjust the control of the 3D printer accordingly (for example by increasing or decreasing heat generation at or near a nozzle 160, 170, or by increasing or decreasing speed of relative movement between the printing head 150 and printing bed 200, or adjusting any of the other process parameters mentioned above as appropriate).

[0158] Figure 6 shows a view of a nozzle 160 with a printing material 10 being extruded therethrough so as to print an object on a printing bed 200. The height, H, and width, W, of the extrudate 10 are marked. Defects in 3D printed parts often stem from inaccuracies in one or both of extrusion height and extrusion width. The height, H, is the thickness of the printed extrudate 10 in a direction perpendicular to the plane of the printing bed 200. In most 3D printer 100 arrangements, this direction is vertical, hence the thickness, H, being referred to as height. The example pictured in Figure 6 is for FFF 3D printing, and the height, H, is the distance between the printing bed 200 and a lower edge of the nozzle 160 - the extrudate 10 fills this vertical space in the example shown. In other scenarios, the height of the printed extrudate will be less than the nozzle-bed spacing. Even in scenarios like that pictured in Figure 6, the height, H, for a subsequent printed layer will be the spacing between the nozzle tip and the top of the previous printed layer as a 3D printed object is built up, so it will be appreciated that bed-nozzle spacing is only a proxy for height in limited circumstances. Nonetheless, height measurement techniques are well known and widely -used, and will not be addressed further herein. The width, W, is the thickness of the printed extrudate 10 in a direction parallel to the plane of the printing bed 200, and perpendicular to a direction of relative movement between the printing head 150 and printing bed 200. In most 3D printer 100 arrangements, this direction is horizontal. The width, W, may be greater than the nozzle outlet 160a diameter. Embodiments of the invention provide an intelligent in situ visual monitoring and measurement system which allows extrusion width to be accurately determined in real-time, and without contacting (and so risking disruption of) the extrudate 10. Although commercial 3D printers can measure the extrusion height using various distance sensors, measuring of the width in real-time remains a challenge, and is addressed herein.

[0159] The 3D printer 100 for use in width determination additionally comprises a camera 154. The camera 154 is arranged to include in its field of view a region in front of a nozzle 160, 170 of the 3D printer 100, where an extruded material will be present. In particular, a camera mount 152 is provided on or as part of the printing head 150, and a camera 154 is mounted on the camera mount 152 in the embodiment being described. As the camera 154 is mounted on the printing head 150, its position with respect to the nozzle 260 is fixed even when / if the printing head moves 150. Mounting the camera 154 on the printing head 150 also allows close proximity to the nozzle 160, so as to obtain a close-up image of the extruded printing material close to the nozzle 160. In the example pictured in Figure 7, the camera 154 is a Raspberry Pi camera - it will be appreciated that any suitable camera 154 may be used in other embodiments. The camera 154 is arranged to provide visual images of the printed extrudate 10. The relatively short vertical distance (e.g. less than 5 cm, and optionally around 3.5 cm) between the camera 154 and nozzle 160 allows the use of a small-sized camera 154 such as the Raspberry Pi v2 with 8 megapixel resolution and 33 mm focal length used in this embodiment. The camera 154 records 60 frames per second with 640x480 pixels resolution. The camera mount 152 of the embodiment shown is adjustable so as to allow the angle of the camera 154 to be adjusted to get a desired view of the printed extrudate 10. In other embodiments, the camera mount 152 may not be adjustable and may determine a single position for the camera 154. The printing head 150 of this embodiment further comprises a light mount 156, and a light 158 mounted on the light mount 156. The light 158 is arranged to illuminate the printed extrudate 10, and the mount 156 is adjustable to allow the light position / angle to be adjusted as appropriate to better-illuminate the extrudate 10. In other embodiments, the light mount 156 may not be adjustable and may determine a single position for the light 158. In other embodiments, no light or light mount may be provided - instead relying on illumination from the environment. In the example pictured in Figures 7 and 8, the 3D printer 100 comprises a shroud 700 surrounding the printing head 150; the shroud 700 may be thought of as forming as part of the printing head 150. The shroud 700 comprises, or has mounted thereon, the camera mount 152 and the light mount 156. The mounts 152, 156 are separately made and mounted on the shroud in the embodiment shown, but either or both may be integrally formed with the shroud 700 in other embodiments. The shroud 700 is rigidly mounted on the printing head 150 such that the shroud has a set position with respect to the nozzle 160, in the embodiment shown. The shroud 700 may be integrally formed with the printing head 150 in other embodiments, however it will be appreciated that a modular design may be favoured in some embodiments for both adjustability and cost-saving reasons. The shroud 700 itself may be 3D-printed to provide a tailored shape with a good fit for a relevant printing head 150 and camera 154, and is 3D printed in the embodiment currently being described - Figure 7 shows the printed shroud 700 from different angles.

[0160] The printer 100 further comprises a processing module 110 arranged to obtain at least one visual image from the camera 154; perform instance segmentation on the visual image to localise and segment the extrudate, so identifying the printed extrudate within the visual image; and perform distance measurement with reference to a calibration standard to determine the extrusion width of the identified printed extrudate. In the embodiment being described, a two-stage instance segmentation is performed, comprising performing image processing on the at least one obtained visual image to localise the extrudate 10 to be measured; segmenting the at least one obtained visual image to distinguish objects from a background and to provide a boundary of detected objects; and combining the results of the separate localisation and segmentation processes to identify the printed extrudate. The processing module 110 of the embodiment shown is arranged to compare the determined extrusion width to an expected extrusion width. The 3D printer 100 comprises a controller 112 arranged to receive a signal from the processing module 110 and to take action if a difference between the two widths exceeds a set tolerance.

[0161] The printer 100 is therefore arranged to be used for monitoring of single line extrusion using machine learning and vision. The width measurement can be used to calibrate the slicing parameters for accurate 3D printing, for which measuring geometry of the printing road is key for detection of anomalies. The printer 100 shown in Figures 7 and 8 compromises a tailor-made printable shroud 700 to hold a camera 154, and software operating in a processing module 110 to apply image processing operations and instance segmentation approaches to the captured images of the extrusion, as described in more detail below. The shroud 700 may be designed to be retrofittable to a standard printing head 150. The shroud 700 of the embodiment being described sits on the printing head 150 and comprises adjustable camera and light mounts 152, 156, which serve to illuminate and monitor the printing roads (i.e. the lines of deposited extrudate 10). In other embodiments, for example in 3D printers 100 built with the monitoring system of this invention in mind, a camera 154 and / or camera mount 152 may be provided as part of the printing head 150 itself, and no shroud may be used.

[0162] Prior attempts to provide in situ monitoring of 3D printing lack the transferability and ease of use. Some studies used a digital microscope, but thus resulted in a bulky structure and small working space. In another approach (Y. Wang, J. Huang, Y. Wang, S. Feng, T. Peng, H. Yang, J. Zou, “A cnn-based adaptive surface monitoring system for fused deposition modelling" , IEEE / ASME Transactions on Mechatronics 25 (5) (2020) 2287-2296), the lack of working space was addressed by moving the camera to an open space around the printing bed, however this modification limited views to the external surface of the print and increased the distance between the camera and the part being printed. Mounting a small camera 154 directly onto the printing head 150 has been found to overcome these challenges whilst still providing accurate measurements using the automated process described below. The use of a shroud 700, and in particular a 3D printed shroud 700, was found to improve compactness and reduce cost of the physical monitoring set-up. A visual in situ monitoring system has therefore been developed to measure width of the printing filament road, and its accuracy thoroughly tested for single-layer printing (i.e. in 2D patterns) - with this successful proof-of- concept study, the approach can be generalised to multi-layer prints by training the models on a more extensive training set including training data with multiple printed layers. The implementation described in detail below uses a visual detection approach based on a two-stage instance segmentation method - it will be appreciated that this example is not intended to be limiting. Each of the segmentation and localisation stages can use a variety of computational approaches including a Gaussian mixture model, colour filters, and deep neural network models. The visual monitoring system discussed herein is arranged to be mounted on a standard 3D printer 100 (the shroud 700 may be retrofitted to a standard printing head 150), and was validated by measurement of printed filament roads of sub -millimetre widths. The accuracy and robustness tests indicated that combinations of deep models for both the segmentation and the localisation have strong performance. Particularly, a fully-connected CNN segmentation model combined with a YOLO object detector was found to be able to measure sub -millimetre extrusion width with 90 pm accuracy at 125 ms speed. This visual monitoring system therefore has the potential to improve the control of 3D printing processes using real-time measurement of printed filament geometry.

[0163] Localisation and segmentation are performed as detailed below to identify the extrudate 10 within the visual image 20 taken by the camera 154. Localisation is used to detect the extrudate 10 to be measured (as there may be multiple objects, optionally including multiple lines of extrudate, within the visual image). Segmentation is used to separate one or more identified objects from a background. Combing the two allows the extrudate 10 to be measured to be identified and separated from the background and any other objects within the visual image 20. The results of the localisation and segmentation are then used to measure the width of the extrudate 10. Figures 9 and 10 illustrate this process. Instance segmentation is a computer vision task that localises instances of objects and predicts their pixel-wise segmentation mask. A measuring module based on instance segmentation was developed to measure the extrudate 10 - the measuring module can use a variety of methods for localisation and segmentation including conventional and state-of-the-art ML models, for example, YOLO or ROI (e g. using a predefined constant box) for localisation, and deep models (such as FCN, Unet, Enet) or more conventional approaches (e.g. GMM, or filtering) for segmentation. Predefined ROI localisation is a manual localisation method which can be used to extract a narrow constant box (e.g. rectangle or other shape, as desired) in front of the nozzle 160 containing the extrudate 10 to be measured. In this method, it is assumed that the extrudate 10 is always confined by this box; however, determining suitable size and place for the box requires trial and error. Various semantic segmentation and localisation combinations were tested and assessed in terms of accuracy, robustness, and speed. Two training datasets were prepared for the two-stage process used - one for training the localisation model and one for training the segmentation model. It will be appreciated that the same model may be used for both localisation and segmentation in a c o mb ined / one- stage approach. The segmentation dataset of the embodiment being described included 1000 images of printed lines with different widths (in the range from 0.4 to 0.8 mm), and various air gaps between them (in the range from 0.3 to 5 mm). The localisation dataset included a subset of the segmentation dataset (400 images), with the 400 images labelled to train the YOLO model used. Extrusions were manually labelled as currently -printing or already-printed roads for YOLO, to provide training data to train the algorithm to recognise the most recent extrudate. The manual labelling specifies the region of interest in each training image provided to the model, and the model then attempts to identify the inherent / robust properties of those regions to distinguish them from the rest of the scene in all circumstances. The initial set of 400 labelled images was augmented to a total number of 1200 images by applying ±15° vertical and horizontal shear, ±13% change in exposure, and adding noise up to ±5% to the pixels. Both datasets were grouped into training and validation sets with an 80-20% split. All images in the segmentation dataset were resized to 128x128 pixels for training of the semantic segmentation CNN models up to 100 epochs with batch size of 16. A check pointer was used to stop the training process once the validation loss ceased to improve by more than 0.0001 in 10 consecutive epochs. The YOLO model was also trained on the localisation dataset for 300 epochs with batch size of 64. For the test dataset, images were captured during the printing process of five lines of extrudate 10 with different widths to evaluate the accuracy of proposed combinations. To print these lines (to form a first test bed), various lengths of filament (11, 8, 6, 4, and 2 mm) were extruded over a 100 mm traveling path. After deposition, width of the lines was measured using KEYENCE VHX-6000 digital microscope at more than ten checkpoints on each line. Table 1, below, provides the average and standard deviation of measured widths for the different length of filament extrusion, with all values in mm. The width measurement results were compared using the symmetric mean absolute percentage error (SMAPE) and mean average accuracy.

[0164] Table 1: Average (mean) and standard deviation (STD) of measured widths for the first test bed SMAPE shows the average deviation percentage of the measured widths from actual width percentages for each line, as follows: where A is the actual extrusion width, M is the measured width, and n is total number measuring points on each printed line. Figure 11 shows the errors of different combinations implemented to measure width of deposited roads in the first test bed. The plots in the left column (a) represent SMAPE of segmentation methods with ROI identification using a constant box. The right column (b) shows SMAPE of segmentation methods in combination with YOLOv5. These results reflect the performance of the entire algorithm including semantic segmentation, localisation, calibration, and determination of the absolute width measurement (e.g. in units of mm), rather than representing the precision of segmentation or localisation methods alone. Analysing the results reveals that even colour filter and GMM perform well, with less than 10% error in measuring “middle range” widths in the range from 0.55 to 0.85 mm (in accordance with the range of widths for which the system was trained and calibrated - it will be appreciated that other systems may be differently trained and calibrated depending on the expected range of extrudate widths. As such, an increased measurement error outside of that range is to be expected, in particular for widths of or below 0.36 mm, with a lower volume of extruded polymer on the road). CNN semantic segmentation models were found to have lower overall errors compared to the more traditional approaches of colour filtering and clustering methods such as GMM. The high accuracy ( >95%) of CNN semantic segmentation models in measuring these middle range widths indicates the power of deep semantic segmentation models to extract masks of the printing roads with sub -millimetre widths. Precision of these models decreases when measuring widths of <0.36 and >1.03 mm, as they are unseen data with respect to the range of widths in the training set. However, these models have still been shown to estimate widths beyond the range of the training data with more than 80% accuracy. Among segmentation models, fully connected architectures were found to provide higher overall accuracy, while U-net and Enet show lower errors for middle range widths, but higher errors outside of the range of the training data. Multiple inaccuracies limit the minimum measurement error that can be achieved by measurement modules as described herein with current technologies - e g. annotation errors in labelling and camera distortion - but a richer dataset still can improve the performance of CNN models. On the other hand, obtaining a general number of GMM clusters or global RGB thresholds for the colour filter are difficult tasks, so favouring deep architectures. A simple comparison between results of YOLO object detection (Figure 11, column (b)) and the ROI constant box approach (Figure 11, column (a)) suggests that the ROI constant box localisation performs slightly better than YOLO. This is because YOLO provides a longer bounding box that cuts the detected extrudate at its corners. For manual localisation to set the constant box for ROI detection, a shorter bounding box was set around the nozzle outlet to capture the whole width of the extruding roads in all circumstances. Despite the better accuracy, ROI constant box localisation lacks the flexibility of YOLO detection, as described in more detail below.

[0165] A second test bed including five roads of 0.8 mm width was printed, sequentially reducing the air-gap between consecutive extruded parts (to 2.25, 1.18, 0.68, 0.18 mm), to measure sensitivity of the approaches tested to interference from other objects within the visual image 20 captured. A 2.25 mm air-gap was present between the first and second lines, and this value was sequentially reduced to reach just 0.18 mm between the fourth and fifth lines. The second test bed was used to measure accuracy with respect to different airgaps between printing roads, to evaluate the vulnerability of localisation methods to visual artifacts as the printed lines get closer together, and the resolution of semantic segmentation models in extracting masks of close roads. The results are shown in Figure 12, which plots the mean average error (MAE) against width of the air-gap between the printing roads. The lines for the use of ROI constant box localisation with any of Enet, FCN, and Unet for segmentation overlap closely, with the Unet-Box combination having a fractionally higher error for air gaps of around 1.18 mm. The lines for the use of YOLO localisation with any of Enet, FCN, and Unet for segmentation again overlap closely, with the combination including Unet again having a fractionally higher error for air gaps of around 1.18 mm. The results indicate that the ROI constant box localisation fails once a part of one or more of the adjacent roads leaks into the constant box (which may also be referred to as an inspection box), which occurred for an air-gap of below 0.68 mm - with the MAE rising sharply for all three plots of combinations including ROI constant box localisation. By contrast, the YOLO detectors showed lower errors, even for air gaps less than 0.5 mm, despite this being unseen data for the model regarding the values of air gaps in the training dataset. In Figure 12, various combinations of segmentation-localisation methods are listed by abbreviation, such that FCN-YOLO stands for FCN semantic segmentation model in combination with YOLO localisation, etc.

[0166] In addition, the speed of various combinations was considered against their measurement error. Figure 13 plots the speed of each method on an NVIDIA Tesla T4 GPU against its mean average error across all measurements in the first test bed. With less than 100 pm error, CNN segmentation models as described herein were found to measure sub -millimetre extrusion widths at least three orders of magnitude faster than methods previously proposed in the literature (see e.g. R. Badarinat, V. Prabhu, "Real-time sensing of output polymer flow temperature and volumetric flowrate in fused filament fabrication process" , Materials 15 (2)).

[0167] Once the extrudate 10 to be measured has been localised and segmented, its absolute width is determined by reference to a calibration standard. The calibration is generally performed in advance of the printing process to be monitored, using a marker 30 of known dimensions, such as that shown in Figure 14. The marker 30 is placed in or near the location in which the extrudate 10 is to be printed, in front of the nozzle 160. A visual image of the marker 30 is then taken by the camera 154. The camera angle and position with respect to the printing head 150 are kept the same for the calibration step and the printing process to be monitored. Figure 14 illustrates a calibration standard, or marker, 30, which has known features for ease of being identified by an image processing system, and known dimensions, alongside a visual image 30a of that marker 30 adjacent to a 3D printing nozzle 160, taken by the camera 154 (which is at an angle to the printing bed 200 on which the marker 30 rests), and a perspective-corrected image 30b of the marker 30 extracted from the camera image 30a. In the embodiment shown, an affine transformation was calculated to map the corners of the marker 30 (as identified by an image processing algorithm) to the corners of a frame with an arbitrary size, and having the known shape of the marker (in this case, square). The width and height of the frame are set to be equal to the known dimensions of the marker. The dimensions of the frame are therefore known both in physical values (from knowledge of the marker 30) and as a number of pixels from the perspective-corrected (transformed) image 30a. A distance calibration parameter can therefore be computed by dividing the physical width of the frame by the number of pixels along the width of the frame. In the real-time application these two parameters (the established affine transformation and the distance calibration parameter) can then be used to correct the perspective of an image of the identified extrudate 10, and map the number of pixels in the extracted mask of the extrudate 10 to a physical value - the extrudate width (optionally in mm). The camera and light mounts 152, 156 are adjustable to enable a user to select a good view of the extrudate 10 as it emerges, or shortly after it has emerged, from the nozzle 160. This adjustment should be performed before any real-time measurements, and also before calibration of the system. The order of operations should be as follows: 1) adjust the camera mount and light to have a good view; 2) calibrate the parameters for that view; 3) keep the position of the camera and lights mounts fixed, and the same as during the calibration step, during all real-time measurements.

[0168] A method 1500 of automatically measuring extrusion width of an extrudate 10 printed by a 3D printer 100 is therefore disclosed, as illustrated in Figure 15. The method 1500 comprises obtaining 1502 at least one visual image 20 of the extrudate 10 whilst a printing process is ongoing. The visual image 20 may include multiple lines of extrudate, and / or one or more other objects (e.g. a portion of the nozzle 160). In the embodiment being described, the visual image 20 is captured by a camera 154 mounted on the printing head 150 through which the extrudate 10 is printed. The camera 154 then provides that data to a processing unit 110. The method 1500 comprises performing 1504 instance segmentation on the visual image 20 using image processing techniques, the instance segmentation comprises localisation of the extrudate, and segmentation to separate the extrudate from the rest of the image. In the example described above, the instance segmentation 1504 is a two-step process comprising performing image processing on the at least one obtained visual image 20 to localise the extrudate 10 to be measured; and performing image processing on the at least one obtained visual image 20 to segment the at least one obtained visual image 20 to distinguish one or more objects (including the extrudate to be measured) from a background and therefore to provide a boundary of the detected object(s). The extrudate localisation information is then combined with the segmentation to identify the printed extrudate 20.

[0169] The method 1500 then comprises performing 1506 distance measurement with reference to a calibration standard to determine the extrusion width of the identified printed extrudate 20. Any suitable calibration approach may be used, including the example with a marker 30 described above. The method 1500 allows extrusion width to be determined in real-time, for example within 0.01 to 0.21 seconds of the measured portion of the extrudate being printed, and optionally within around 0.125 seconds of the measured portion of the extrudate being printed. Embodiments were found to allow this high speed of image processing and width determination whilst maintaining high robustness against visual artifacts. For example, a test system was shown to successfully measures each 10 mm length of a printed extrusion in 125 ms.

[0170] In various embodiments, the method 1500 further compromises taking action based on the determined width measurement - for example, the processing module 110 may compare the measured width to an expected width, and send instructions to the controller 112 of the 3D printer 100 if the measured width is not within tolerances of the expected width. The controller 112 may adjust the control of the 3D printer accordingly to correct the width, or may terminate the printing process.

[0171] Figures 9 and 10 illustrate implementation of the method 1500. Figure 9 shows an input visual image 20, including an image of an extrudate to be measured, being fed to a processing module 110. Within the processing module 110, localisation and segmentation are performed. Figure 9 shows various possible combinations of methodologies for a two-step approach; it will be appreciated that these are listed by way of example only. The localisation method detects / identifies the printing line - i.e. the section of extrudate currently being printed, and which is to be measured. This is shown by way of a second image 20a, with the printing line marked with a lighter-coloured box, and other (earlier-printed) lines of extrudate marked with darker-coloured boxes. In the embodiment being described, the localiser is trained to detect all lines of extrudate in a given frame and label them as either “currently -printing” or “printed”. The algorithm then only uses the “currently -printing” detection for further frame processing to extract the width. The segmentation method separates objects from a background, so providing a mask of the printing scene contained within the input image 20, showing the objects separated from the background. The two outputs 20a, 20b are then combined to provide a mask 20c showing the printing line / extrudate of interest only, with the rest of the background and any other detected objects removed. That output is then used for the width measurement. If necessary, a perspective transformation (e.g. an affine transformation) is performed to correct for the angle between the camera 154 and the printed extrudate 10, so providing a corrected mask 20d and / or image 20e. The mask 20c in such embodiments may be referred to as a Transformed Mask of Printing Road (TMOPR) - i.e. a mask of the extrudate that has been transformed by the perspectivecorrection transformation. A distance measurement across the width of the extrudate 10 is then extracted, by reference to the earlier calibration of the system. To measure the width of extrudate, the edges of the extracted mask, or TMOPR 20c, are approximated by a small rectangle, so linearising each edge of the extrudate in the TMOPR 20c with a straight line. This may result in an inaccurate measurement, as these edges are often not straight even over the relatively short distance captured in each frame, for example due to variation of the printing conditions. Therefore, a piece-wise linearisation approach was used to better approximate the extrudate’s edges by slicing the mask (TMOPR) 20c into several vertical segments (in the orientation shown) and approximating a rectangle for each segment separately. The width of the extrudate 10 in each frame is calculated as an average over the widths of all rectangles. In the segmented images 20f shown, twelve segments / rectangles per frame were used to approximate the edges of the extrudate. As the number of segments increases, the accuracy of measurement improves but there is a trade-off in terms of computational cost - having more than twelve segments in a 10 mm section was not found to provide a sufficient improvement in accuracy to merit the additional processing power. This segmentation is shown by the small boxes / rectangles visible along the length of the extrudate 10 in the segmented images 20f.

[0172] Figure 10 provides a table of images demonstrating the same process for localisation and segmentation, for printing scenes of increasing complexity (in particular, increasing numbers of similar visible objects within the visual image 20). Figure 10 also demonstrates the order of computational process for the measurement algorithm - from image capture 20, through localisation 20a and segmentation 20b, and onto a perspective correction 20c and measurement 20f. It will be appreciated that the visual representations of the masks etc. may be provided only for use by a human user, and that the computational system may use the raw data of the masks / images without ever generating an actual visible image.

[0173] Among all CNN combinations tested, FCN segmentation with YOLO object detection showed better accuracy in the tests performed even within relatively small air gaps. With prediction times of just 125 ms per frame, the FCN-YOLO method can support printing speeds of up to 83.3 mm / s. In the embodiment being described, over 12 different measurements are taken from a 10 mm length of extrudate 10 in each frame, and these are averaged, so assuming an at least substantially constant width along the 10 mm section of extrudate. Faster combinations based on predefined ROI (e.g. constant box) detection also are available for controlled condition trials, for example in real-time calibration of 3D printers 100. These models can measure the extrusion width at printing speeds of up to 100 mm / s, but may struggle in real-world applications (e.g. with multiple close lines of extrudate). A low-cost and compact visual set-up for in situ monitoring of 3D printing roads is therefore disclosed - whilst the examples described in detail relate to fused filament fabrication (FFF) processes, it will be appreciated that the same principles can be applied to other 3D printing modes, with different training data and potentially a different camera 154 and / or camera angle, as appropriate.

[0174] The feasibility of using two-stage instance segmentation to measure extrusion width in real-time was assessed. For the segmentation and localisation tasks, multiple models were tested including conventional ML and CNN, showing that a variety of techniques can be used, and in varying combinations, to achieve the principles disclosed herein. The experiments on 3D printed lines with various extrusion widths and air-gaps demonstrated that an FCN-YOLO combination achieved the highest performance out of the methodologies tested, giving a measurement resolution of 90 pm, a processing time of 125 ms, and showing robustness against small air gaps (so processing around 8 frames / second with high accuracy and reliability). The in situ measurement system disclosed herein therefore opens up opportunities for real-time anomaly detection and quality control of the 3D printing at the level of printing roads (i.e. the level of individual extruded strands). It will be appreciated that there are various instance segmentation approaches, any of which may be implemented for the method described herein - whilst the examples described in detail herein focus on a two-stage approach, this is not intended to be limiting. Selection of an optimal method is a trade-off between speed and accuracy: as the number of stages increases, the accuracy generally improves while the computational cost increases. With current technology, two-stage methods have performed better for small object identification. Moreover, the modular architecture of two- or multi-stage approaches allows for easy adaption of the model. The performance of the two-stage model strongly relies on accuracy of localisation, which is generally an easier task than segmentation. Therefore, a well-performing and fast object detection model with pre-trained weights was used for detection of many objects (YOLO) and used transfer learning to teach the network to detect the extrudate 10 of interest for this work. The strong performance of the localiser allowed the use of simpler segmentation models without overly compromising accuracy.

[0175] Turning back to physical features of the printer 100 arranged to implement this method 1500, Figure 16A shows an example of the shroud 700. In Figures 16B and 16C, this shroud 700 is shown mounted on a printing head 150, and with a camera mount 152, camera 154, and light mount 156 mounted onto the shroud 700. The shroud 700 has a generally rectangular cross-section. In the embodiment shown, the shroud 700 has two camera mounts 152 mounted thereon - a first camera mount 152a mounted on a first side of the shroud 700, and a second camera mount (made of two parts, 152b, 152c) mounted on a second side of the shroud 700, the second side of the shroud 700 is perpendicular to the first. Each camera mount 152 has a camera 154a, 154b mounted on it - the shroud 700 therefore has two cameras 154. The printing head 150 of this embodiment has two nozzles 160, 170, and each nozzle has a dedicated camera 154a, 154b. In other embodiments, the printing head 150 may have more or fewer cameras 154 than it has nozzles 160, 170 - for example, with one (optionally movable) camera 154 covering multiple nozzles, or a single nozzle having multiple cameras to obtain different views and / or different resolutions (e.g. for different printing modes of the same nozzle). In the embodiment shown, each camera mount 152a, 152b-c is arranged to be adjustable by rotation around at least one axis, so as to allow camera angle to be adjusted. The camera 154a, 154b are both the same type of camera in this embodiment, but different kinds of camera may be used in some embodiments, optionally for different nozzles 160, 170 - for example, a higher-resolution camera may be used for a nozzle 170 intended for electro-spraying or electrospinning, due to the extrudate generally being thinner than for FFF. The camera mount components 152a, 152b, 152c are shown individually in Figure 17. Whilst two different designs of camera mount 152 are shown in the embodiment being described, this is simply due to space limitations of the specific printer 100 in use, and it will be appreciated that the same kind of mount could be used for each of multiple cameras in other embodiments. In the embodiment shown, the shroud 700 similarly has two light mounts 156 mounted thereon - a first light mount 156a for the first camera 154a, and located near to that camera 154a, and a second light mount 156b for the second camera 154b, and located near to that camera 154b. two lights 158 are provided in this embodiment - one for each camera 154. The light mounts 156a, 156b are shown individually in Figure 18. Again, whilst these mounts 156a, 156b differ from each other, the same kind of mount could be used for each of multiple lights in other embodiments - design specifics are to be selected as appropriate for a given printer 100. It will be appreciated that the mounts 152, 156 may take any suitable shape, which may depend on the type of camera 154 and light 158 as well as printing head design and nozzle arrangement. The mounts 152, 156 generally have complex shapes with strict tolerances, and may be made by 3D printing. The shroud 700 and mount components 152, 156 may be 3D printed (generally by FFF) and assembled as a modular apparatus, so facilitating manufacturing of parts to retrofit an extant 3D printer 100 so as to perform the measurement technique disclosed herein. The use of multiple cameras 154 may be particularly beneficial when the printer 100 is a multi-modal printer 100 capable of printing in both an FFF mode (for which the extrudate 10 is generally thicker) and an eHD mode (for which the extrudate 10 is generally thinner, and which may be performed by a different nozzle from the FFF mode), and / or capable of printing via multiple different nozzles 160, 170 (which may be spaced apart such that use of a different camera for each is suitable). Even when only a single nozzle is used, providing two or more cameras 154 to monitor the nozzle from different angles, so providing multiple images of the extrudate 10, may help to improve accuracy of the measurement. Multiple lights 158 located in different positions may similarly be beneficial even for a single nozzle to provide better illumination and avoid shadows.

[0176] Figure 19 shows the printing head 150, including the shroud 700, in context in a portion of a 3D printer 100, showing part of the printer frame 101 which supports various components of the 3D printer 100, including the printing head 150. The printing head 150 is mounted on the frame 101 on a movement carriage arranged to allow vertical motion, and optionally also horizontal motion, of the printing head 150 along the frame 101. Multiple fans 151 are mounted on or around the printing head 150 - the fans are controllable so as to reduce temperature in or around the nozzle 160 when desired.

[0177] The camera 154a mounted on the printing head 150 may be described as a front camera in the embodiment being described, as it is mounted in front of the nozzle 160 and arranged to take a close-up picture of the emerging extrudate 10. A side camera may also be provided in some embodiments, further from the nozzle 160 and aimed more generally at the printing area, to provide an overview of the printing process. Such a side camera may be separately mounted on the printing frame 101 - not on the printing head 150 - such that it can be moved independently from the printing head 150. No side camera may be provided in other embodiments, or a differently-positioned camera may be used instead, optionally being movable such that its position can be changed when a printing mode changes. One or more other sensors may also be provided - mounted on the printing head 150, printing bed 200, and / or on the printer frame 101 or other parts of the 3D printer 100 in various embodiments, forming parts of a monitoring system 111.

[0178] The ability to switch between typical, extrusion-based, material printing (FFF) and eHD printing requires an electromechanical system that can rapidly adapt to the application of the high applied voltage in eHD modes. This necessitates careful design of the mechanical system. In particular, an insulating printing bed apparatus 200 was designed to replace a “traditional” printing bed - the insulating printing bed apparatus 200 is arranged to be electrified by high DC voltage with minimal voltage and current leakage, and can act as both a “traditional” printing bed for FFF printing and as a collector for eHD printing. The specific example of the insulating printing bed apparatus 200 described in detail herein was designed to have printable parts (printable by an FFF 3D printer), and to be modular, so as to provide ease of manufacture at a relatively low cost. This specific printing bed apparatus 200 was also designed to be retrofittable to a standard FFF 3D printer, re-using components of the original 3D printer where applicable to “upgrade” it for use as a multi-modal 3D printer 100. It will be appreciated that different fabrication methods may be used in other embodiments, in particular where a multi-modal 3D printer 100 is being newly built, rather than made as an upgrade to an extant printer. The insulating printing bed apparatus 200 is shown in Figure 20 and comprises four main components - a conductive plate 202, an electrically -insulated layer 204, an electrical insulation frame 206, and a spacer 208. The conductive plate 202, which is a rectangular plate of 3 mm thick aluminium in the example being described, is arranged to be electrified when the printer 100 is operated in the eHD 3D printing mode. When the printer 100 is operated in an FFF mode, no electrification is needed and the conductive plate 202 instead simply provides a flat surface onto which the extrudate 10 can be printed, acting as a standard printing bed. The electrically-insulated layer 204 lies below (in the orientation shown) the conductive plate 202 and parallel to the conductive plate 202 In use, the conductive plate 202 lies between the electrically-insulated layer 204 and the printing head 150. The electrically-insulated layer 204 takes the form of a flat plate, like the conductive plate 202, and is of at least substantially the same face shape and area as the conductive plate 202 in the embodiment shown. In embodiments such as that shown in which an FFF 3D printer is being retrofitted with additional components to function for eHD, the electrically-insulated layer 204 may be provided by an original printing bed of the 3D printer. In the embodiment being described, the extant printing bed had electrically insulating material added to it - more specifically in the form of two layers of mica plastic film with dielectric strength of 25 kV / mm, one on each face of the original printing bed - to form the electrically-insulated layer 204. If building a 3D printer afresh rather than making parts for a retrofit, the insulated layer 204 may be provided by an insulating sheet - e.g. a 3 mm thick PVC plastic sheet. Keeping the sheet relatively thin may allow for a small amount of bending, e g. to facilitate fitting within the frame 206 described below,

[0179] The electrical insulation frame 206, which is 3D printed in the example being described, surrounds the edges of the conductive plate 202 and electrically isolates the conductive plate 202 from the printer frame 101 and the printer’s motion system. The electrical insulation frame 206 also serves to mount the conductive plate 202 on the insulating layer 204. A thickness of the electrical insulation frame 206 spaces the conductive plate from the insulating layer. The electrical insulation frame 206 has a lower lip 206a arranged to hold the insulating layer 204, such that the insulating layer 204 is held in a position near a lower surface of the electrical insulation frame 206. A typical spacing between the conductive plate 202 and the insulating layer 204 may be between 5 mm and 25 mm, and typically around 10 mm. The electrical insulation frame 206 is made of a resiliently deformable material in the embodiment being described, such that it can be stretched / bent out of shape to fit the conductive plate 202 and the insulating layer 204 within it, and then returns to / towards its original shape when the force is released, so gripping the two layers 202, 204. The electrical insulation frame 206 is described as a “frame” because it surrounds the edges of both layers 202, 204 but does not cover their surfaces, so framing them.

[0180] The spacer 208 is mounted on the insulating layer 204, on the opposite side of the insulating layer 204 from the conductive plate 202. The spacer 208 is also made by 3D printing in the embodiment being described. The spacer 208 serves to distance the conductive plate 202, which may be described as forming an electrifiable printing bed, from a motion carriage 120 of the 3D printer 100, along a vertical axis (in the embodiment shown). The spacer 208 of the embodiment being described comprises at least one bed-levelling adjustor 209. Figure 21 shows the spacer 208 mounted between the insulating layer 204 and the motion carriage 120, without the conductive plate 202 and insulating frame 206. The use of a modular compliant structure, with the insulating frame 206 being arranged to grip the conductive plate 202 and insulating layer 204 means that the conductive plate 202 can be mounted on a motion system 120 without using metal screws. Figure 22 shows a cross-section of the insulating printing bed apparatus 200, showing the use of the printed compliant insulating frame 206 to eliminate the need for screws in the assembly 200. In the embodiment shown, the spacer 208 comprises four parts - 208a-d - one located in each corner region of the insulating layer 204, as shown in Figure 23. The spacer 208 is arranged to distance the conductive plate 202 from the motion carriage 120 by a distance of around 45 mm in the example pictured, although this may vary in other embodiments. Considering the example of Polylactic acid, PLA, which is a thermoplastic material sometimes used as a filament in 3D printing and which could be used to make the spacer 208, its reported dielectric strength is around 33-50 kV / mm - even a 1 mm tall spacer 208 should therefore suffice for a maximum applied voltage of 20 kV. However, it will be appreciated that a thicker spacer will generally reduce current leakage.

[0181] If building a printer 100 afresh rather than retrofitting components to an extant motion carriage 120, a motion carriage plate may be made from an insulating material, e.g. by custom machining of a thick (e g. 2 cm thick) insulating plastic sheet (e.g. PVC, or PTFE, or a synthetic resin). In such embodiments, a shorter spacer 208 may suffice as the motion carriage plate provides some additional insulation.

[0182] Each part 208a-d of the spacer 208 has a part of the bed-levelling adjustor 209 mounted thereon. In the embodiment shown, the bed-levelling adjustor 209 comprises multiple spring and wheel mechanisms for bed levelling. One spring and wheel mechanism is provided in each corner region of the insulating layer 204 / on each part of the spacer 208a-d. Each spring 209a is adjustable in length by turning the respective wheel 209b. Any suitable bed-levelling adjustor 209 may be used in other embodiments - the spring and wheel mechanism is provided by way of non-limiting example.

[0183] Figure 24 shows the assembly 200 retrofitted to a commercial 3D printer (a Creality Ender3). In these photographs, the electrical contact 202a used to electrify the conductive plate 202 is also shown. Using this bed assembly 200, current leakage during movement of the bed was found to be significantly reduced as compared to when using a standard / traditional 3D printing bed.

[0184] Figure 25 illustrates a method 2500 of retrofitting a printing bed of an FFF 3D printer to allow usage of the printing bed in at least one electrohydrodynamic - eHD - 3D printing mode (optionally as well as in one or more FFF printing modes). The method 2500 comprises coating 2502 each face of the printing bed with an insulating material to form an electrically-insulated layer 204, e g. by applying a mica plastic film to each face of the printing bed. Any other suitable material may be used in other embodiments, and the insulating material may be applied as a liquid (which then dries and / or is cured), or as a solid film (which may be connected to the printing bed using an adhesive or any suitable fastening means). In some embodiments, solid insulating layers may simply be placed on either side of the printing bed, and held in place by the electrical insulation frame 206 described elsewhere herein. The method 2500 comprises mounting 2504 a conductive plate 202 on and parallel to the electrically -insulated layer 204 using an electrical insulation frame 206 which surrounds the edges of the conductive plate 202. A thickness of the electrical insulation frame 206 spaces the conductive plate 202 from the electrically-insulated layer 204 - the electrical insulation frame 206 may comprise a first indentation or ledge arranged to receive the edges of the conductive plate 202, and a second indentation or ledge arranged to receive the edges of the electrically -insulated layer 204 - a thickness of the electrical insulation frame 206 between those two ledges may define the spacing between the two plates 202, 204. This step 2504 may be performed by deforming a specially-shaped resilient electrical insulation frame 206 to insert the two plates 202, 204 therewithin. The method 2500 further comprises mounting 2504 mounting a spacer 208 on the opposite side of the electrically- insulated layer 204 from the conductive plate 202. The spacer 208 is arranged to distance the conductive plate 202 from a motion carriage 120 of the original FFF 3D printer 100. The electrical insulation frame 206 is arranged to electrically isolate the conductive plate 202 from the motion carriage 120. The spacer 208 comprises at least one bed-levelling adjustor 209, and the method 2500 may further comprise adjusting the at least one bed-levelling adjustor 209 to get the conductive plate 202 exactly where it is wanted for a printing process.

[0185] Aside from the insulated bed assembly 200, a multi-modal printer 100 requires a printing head 150 designed to deposit materials in different states (e.g. melt and solution) and / or by different processes (e g. FFF and eHD, and / or different forms of eHD). The printing head 150 described in detail below comprises two nozzles - one 170 for solution deposition, and one 160 for melt deposition. A typical hot-end extruder used for FFF filament printing can also be used to print the melt, and so is selected for use as the melt nozzle 160 in the embodiment being described to reduce complexity, as the same nozzle 160 can be used for both FFF printing and melt eHD. Nozzles 160, 170 used for eHD processes must be electrically-grounded. For the melt nozzle 160, the whole nozzle 160 is often made of metal even for FFF printing processes, so grounding this nozzle 160 is simply a matter of connecting a ground connection to the nozzle 160 itself, generally in the region of the nozzle tip or outlet 160a.

[0186] Solution nozzles 170 are generally more complex in design, often comprising a central needle 172 (or other spinneret 172) as shown in Figure 26. Spinneret sizes and designs can vary between nozzles 170, so an adjustable spinneret grounding apparatus 174 is desirable if it is to be used with multiple different solution nozzles. Solution nozzles 170 generally comprise, or have mounted thereon, a feeder cap (also referred to as a feeder tip) 171 which connects a feeding tube to the spinneret 172. A syringe and plunger mechanism 173, often including an electronically-controllable pump (or any suitable means of moving a syringe plunger), are generally located at the other end of the feeding tube, and arranged to push the fluid through the tube when desired, at a desired rate.

[0187] The spinneret grounding apparatus 174 shown in Figure 27 comprises two conductive contacts 176a, 176b which are biased towards each other. The conductive contacts 176 are arranged on opposing sides of the spinneret 172, which in this case is a needle 172 (so the grounding apparatus 174 may be referred to as a needle grounding apparatus 174), from each other such that they are biased to grip the needle 172 therebetween. The biasing direction of the contacts 176 is perpendicular to the length of the needle 172. The faces of the contacts 176 nearest the needle 172 may be flat, or may comprise a groove parallel to the needle’s length and sized to partially accommodate and grip the needle 172. At least one of the contacts 176, and preferably each conductive contact 176, is electrically grounded. The needle grounding apparatus 174 is mounted on the printing head 150 so as to move with the nozzle 170. As the needle grounding apparatus 174 contacts the needle 172 with moving contacts, it can apply ground polarity to needles of different sizes (as the contacts can move closer together or further apart), and indeed to spinnerets 172 of different types, whilst also keeping the needle or other spinneret 172 steady.

[0188] In the embodiment shown in Figure 27, the apparatus 174 is symmetrical - both contacts 176 are mobile, sliding, contacts biased by springs. In other embodiments, one contact 176a may have a fixed position, and the other contact 176b may be biased towards it - the relative movement still allows needles of different sizes, or spinnerets of different designs, to be gripped. However, use of two mobile contacts 176a, 176b may be favoured to provide more flexibility in spinneret width and positioning. The mechanical linkages 176, 177, 178 automatically grasp spinnerets 172 of various sizes and provide a ground connection. In the embodiment shown in Figure 27, the needle grounding apparatus 174 applies ground voltage onto the needle 172 using two single-acting spring return cylinders 178a, 178b. The contacts 176a, b move along the respective cylinder 178a, 178b under the influence of the springs 177a, 177b. The contacts 176a, b each comprise a cylindrical portion arranged to be a sliding fit within the respective spring return cylinders 178a, b, so limiting movement to a single axis. It will be appreciated that a different cross-sectional shape may be used in place of a circle in some embodiments, so the components may not be cylindrical - the same principles apply. Each cylinder 178 is made of or comprises a conductive sheet or strip - e.g. copper foil - arranged to remain in sliding contact with the conductive contacts 176 as the contacts move. A stationary connection from the conductive part of each cylinder 178 is made to the ground polarity of a power supply. In some embodiments, the cylinders 178 may be made entirely of an electrically -conducting material, although an insulating outer layer may be desirable in many implementations. In other embodiments, instead of having a cylinder 178 surrounding the spring 177 as shown in Figure 27, an end of the spring 177 furthest from the needle 172 may be mounted on a cylinder 178 or peg which fits within the spring, and the spring may be sufficiently rigid to remain aligned with that cylinder along its length, optionally in combination with sliding support from an aperture in a wall through which the contact 176 slides. It will be appreciated that any suitable support structure may be provided, and that the electrical contact(s) to ground may be adjusted as applicable. Figure 28 shows views of a printing head 150 with a solution nozzle 170 adjacent a melt nozzle 160 with (c) and without (a and b) a needle grounding apparatus 174 generally as described above mounted on the printing head 150. The embodiment of Figure 28 uses a supporting peg 178 and an aperture in a supporting wall to support each spring 177 rather than a cylindrical container as shown in Figure 27.

[0189] As shown in Figure 28, an electrically-insulating wall 179 may be provided between the nozzles 160, 170, to reduce or avoid charge transfer. The electrically-insulating wall 179 is provided by a portion of the support structure for the needle grounding apparatus 174 in the example shown, but may be provided separately in other embodiments. The electrically-insulating wall 179 takes the form of a straight, planar, wall between the two nozzles 160, 170 in the example shown, but may take a different shape (e.g. a cylinder surrounding at least part of the needle 172) in other embodiments. Since the effective surface areas of the nozzles 160, 170 are generally not equal, the height and thickness of the wall 179 can be extended as appropriate to avoid charge transfer between the two nozzles, especially at relatively high applied voltages (e.g. > 15 kV). The effective surface area of each nozzle can be computed by modelling them as cylinders. The surface area of the cylinder is equal to:

[0190] A = 2nr2+ 2nrh where r and h are the cylinder radius and height respectively. Therefore, the charge density on the nozzle can be computed as: where the q is the charge on the surface of nozzle. The strength of electrical field around the nozzle can be calculated as: where the is the permittivity of the air. Since the permittivity is same for both nozzles, and with the assumption that the charge is equal on both nozzles:

[0191] Therefore the gradient of the electrical field between the nozzles 160, 170 is proportional to the difference between their effective surface areas. According to the cylindrical model of the nozzles, the effective surface area is highly affected by nozzle’s radius. Therefore, a high difference between nozzle radiii can lead to a corona discharge between the nozzles 160, 170. As the needle 172 of the solution nozzle 170 generally provides a smaller nozzle radius than the nozzle outlet 160a of the melt nozzle 160, the insulating wall 179 may be necessary to prevent corona discharge when the nozzles 160, 170 are in close proximity - as they generally will be when both mounted on a single printing head 150.

[0192] Figures 26 and 28 to 31 illustrate a first design of the needle grounding apparatus 174, with two springs

[0193] 177 and two moving contacts 176 as described above. A support structure 178, 179 of the needle grounding apparatus 174 is shown alone in part (b) of Figure 30, and with the contacts 176 and springs 177 mounted thereon in part (a) of Figure 30. In this embodiment, the support structure 178, provides a peg 178a that slots inside, and serves to firmly locate one (outer) end of the respective spring 177, and a wall portion surrounding an inner end (nearer the needle 172 in use) of the respective spring 177. The contact 176 is mounted on the inner end of its respective spring 177, and passes through an aperture in the wall portion of the support structure 178 which surrounds the spring. The contact 176 is arranged to be a sliding fit in the aperture, so keeping the inner end of the respective spring 177 in line. Each contact 176 comprises a peg 176x arranged to slot into the inner end of the spring, so as to mount the contact 176 on the spring 177, and a lip 176y arranged to sit between the spring 177 and the wall portion of the support structure 178 and to be too wide to fit through the aperture in the wall portion of the support structure 178, so ensuring that the contact 176 cannot fall out of the support structure 178 when no needle 172 is present. A ridge 176z is provided along at least one side of the contact 176, so disrupting the rotational symmetry of the generally cylindrical contact 176. The aperture in the wall portion of the support structure 178 is shaped to have a corresponding indentation, so preventing rotation of the contact 176 in use. It will be appreciated that many different support structure and / or contact shapes may be used in other embodiments, and that a different or additional biasing member may be used in place of, or as well as, a helical spring 177. The support structure

[0194] 178 extends to a greater height to provide an insulating wall 179 between the nozzles 160, 170 in use. The insulating wall 179 is located outward of the outer end of the respective spring 177, and supports the peg 178a in the example pictured. The needle grounding apparatus 174 is symmetrical in this embodiment - having a plane of mirror symmetry between the contacts 176 - and may therefore be used whether the solution nozzle 170 is on the left or right of a melt nozzle 160. Two insulating walls 179 are therefore provided. A single, continuous, wall around three or more sides of the spinneret 172 may be provided in other embodiments.

[0195] Figure 32 illustrates a printing head 150 substantially like that shown in Figure 26, but with an alternative spinneret grounding mechanism 1174. The alternative spinneret grounding mechanism 1174 again incorporates two contacts 1176 each biased towards the needle (or other spinneret) 172, by a biasing member such as a spring 1177. Figure 33 provides three perspective views of the same printing head 150 and spinneret grounding mechanism 1174, and the components of the spinneret grounding mechanism 1174 are shown in more detail in Figures 34 and 35 The contacts 1176 are made of copper sheet foil in the embodiment being described, and more generally may be made of, or covered with, any thin metal sheet, and are electrically connected to ground.

[0196] This spinneret grounding mechanism 1174 comprises a support structure 1178 and two conductive contacts 1176 suspended therefrom. Each contact 1176 is substantially planar, with a flat edge arranged to contact the needle 172 in use. Each contact 1176a,b and has two rotational connection points 1176x, 1176y, each arranged to be connected to a mobile connective member 1177. Two mobile connective members 1177, which may be referred to as connector arms, are arranged to connect each contact 1176 to the support structure 1178. A pair of connective members 1177a, 1177b is provided for each contact 1176a, 1176b. The first and second connective members 1177a, 1177b of each pair each comprise a round peg 1177x, 1177y arranged to fit rotatingly into a respective rotational connection point 1176x, 1176y of the contact 1176, each connection point taking the form of a round indentation, hole or aperture 1176x, 1176y through or into the respective contact 1176 in the example shown. A lower end of each connective member 1177a, 1177b is therefore rotatably connected to a contact 1176a, 1176b. The upper end of each connective member 1177a, 1177b is movably mounted on the support structure 1178, each having a round peg arranged to fit into an aperture through, or cavity in a wall of, the support structure 1178. The outer connective members 1177a (further from the needle) each comprise a hooked arm portion 1177z arranged to pass through an aperture in an upper surface of the support structure 1178. A spring 1177c extends between the hooked arm portion 1177z of each inner connective member 1177a and a connector 1178a of the support structure 1178, biasing the hooked arm portion 1177z outwards / towards the connector 1178a. Pulling the upper end of the connective member 1177a outward causes the lower end of the connective member 1177a to move towards the spinneret 172, so pressing the contact 1176 against the spinneret 172, so providing grounding of the spinneret 172. The connective members 1177 are therefore arranged to allow the contacts 1176 to move towards each other and apart whilst keeping the flat edge of the contact at least substantially parallel to the needle 172.

[0197] Feeder cap sizes can also vary between solution nozzles 170. A modular nozzle system that can hold feeder caps 171 of different sizes and apply a ground voltage to spinnerets 172 of varying sizes was therefore developed - as for components of the bed assembly 200, 3D printability of parts was favoured for ease of manufacture by 3D printer users, so facilitating retrofitting of the apparatus to extant 3D printers 100.

[0198] The solution nozzle 170 of the embodiment being described also comprises a cap holder 175 as shown in Figure 37, which provides a side view and four perspective views of the same cap holder 175. The cap holder 175 was 3D printed using a compliant material, to form a holder 175 with a precise tailored shape, and the ability to deform so as to firmly grip appropriately sized and shaped items fitted therewithin. In particular, the cap holder 175 comprises a block 175a with a cylindrical aperture or passage 175b therethrough - the cylindrical passage 175b is arranged to hold a feeder capl71 mounted therein, the feeder cap 171 being arranged to provide the solution to the needle 172 for printing. The block 175a fully surrounds the feeder cap 171 in an upper portion, so gripping the feeder cap 171 firmly. In a lower portion, a front half of the block 175a is removed, so leaving a rear part of the passage 175b to accommodate and locate the feeder cap 171 and providing a view of the feeder cap 171 and the top of the needle 172 where it extends from the rest of the solution nozzle 170 from the front. Two connectors 175c are also provided; in the embodiment shown, these connectors 175c are simply apertures for screws to pass through, although different connectors may be provided in other embodiments. The connectors 175c are offset from an axis of the passage 175b so as not to interfere with the feeder cap 171.

[0199] It will be appreciated that the particular embodiments described in detail herein are provided by way of example only, and are not intended to limit the scope of the claims. Many possible variations will be apparent to the skilled person on reading this disclosure.

Claims

CLAIMS1. A method of generating process parameters for a 3D printing process, the method comprising: obtaining vibrational spectroscopic data of a material to be printed; providing the spectroscopic data of the material to be printed to a machine learning model, the machine learning model having been trained on a dataset of vibrational spectroscopic data of sample materials to predict at least one of the material of a sample, and one or more thermophysical properties of a sample from spectroscopic data; predicting at least one of the identity of the material to be printed, and one or more thermophysical properties of the material to be printed using the machine learning model; and determining one or more process parameters for the printing process using the material to be printed based on the at least one prediction regarding the material to be printed.

2. The method of Claim 1, further comprising controlling the printing process in accordance with the one or more determined process parameters.

3. The method of Claim 1 or Claim 2, wherein the spectroscopic data are or comprise IR spectra, and optionally are or comprise at least one of Fourier Transform Infrared - FTIR - spectra, and Near Infrared - NIR - spectra.

4. The method of any preceding claim wherein the printing process is a fused filament fabrication - FFF - 3D printing process, or an electro -hydro dynamic - eHD - printing process.

5. The method of any preceding claim wherein the printing process is a fused filament fabrication - FFF- 3D printing process, and wherein the at least one process parameter is or comprises one or more of:(i) filament diameter;(ii) nozzle temperature;(iii)feed rate;(iv) printing speed;(v) printing shear force;(vi) printing bed temperature; and(vii) hot-end temperature.

6. The method of any preceding claim, wherein the printing process is a material jetting 3D printing process, and in particular is an eHD printing process, and wherein the at least one process parameter is or comprises one or more of:(i) mode of operation, for example whether to use melt or solution eHD;(ii) voltage;(iii) nozzle temperature;(iv) nozzle height;(v) flowrate of melt or solution;(vi) Taylor cone modality;(vii) temperature of an eHD chamber;(viii) humidity of an eHD chamber; and (ix) printing speed.

7. The method of any preceding claim, wherein the printing process is a multi-modal printing process, and wherein the at least one process parameter is or comprises a selection of a printing mode, the selection optionally being made between FFF 3D printing, melt electrospinning, and solution electrospinning.

8. The method of Claim 7, wherein the method further comprises detecting a current flowing within printing control circuitry of a 3D printer implementing the method, inferring in which printing mode the process is currently operating from the current, and stopping the printing process if there is a mis-match between the inferred printing mode and an intended printing mode.

9. The method of any preceding claim, wherein the step of obtaining the spectroscopic data comprises obtaining data from an IR spectrometer or Raman spectrometer arranged in-line with the printing process, such that an IR spectrum or Raman spectrum is generated for the material to be printed as the material moves towards a printing nozzle.

10. The method of any preceding claim, wherein the method is arranged to predict at least one of the glass transition temperature of the material and the melt temperature of the material from its spectroscopic data.

11. The method of any preceding claim, wherein the determining one or more process parameters comprises use of both the at least one prediction regarding the material to be printed and knowledge of one or more characteristics of the part to be printed.

12. The method of Claim 2 or any claim dependent thereon, wherein the spectroscopic data of a material currently being printed by a 3D printer is obtained and used to determine the process parameters, and wherein the method further comprises: automatically measuring extrusion width of an extrudate being printed by the 3D printer, by: obtaining at least one visual image of the extrudate whilst a printing process is ongoing; performing instance segmentation on the visual image to localise and segment the extrudate; and performing distance measurement with reference to a calibration standard to determine the extrusion width of the identified printed extrudate; comparing the determined extrusion width to an expected extrusion width; and adjusting one or more process parameters if a difference between the determined extrusion width and the expected extrusion width exceeds a set tolerance13. A 3D printer with automatic printing process parameter determination, the 3D printer comprising: a spectrometer arranged to obtain vibrational spectroscopic data of a printing material within the 3D printer; a processing module, the processing module being arranged to: use a machine learning model to predict at least one of the identity of the material to be printed, and one or more thermophysical properties of the material to be printed, from thespectroscopic data, the machine learning model having been trained on a dataset of spectroscopic data of sample materials; and determine one or more process parameters for the printing process using the material to be printed based on the at least one prediction; and a controller, the controller being arranged to control the 3D printer in accordance with the one or more determined process parameters.

14. The 3D printer of Claim 13, wherein the spectrometer is an IR spectrometer and comprises an Attenuated Total Reflection - ATR - probe located so as to take a spectrum of the printing material either:(i) in a storage location of the printing material within the 3D printer; or(ii) at a location along a travel path of the printing material between the storage location and a printing head of the 3D printer.

15. The 3D printer of Claim 13 or Claim 14, further comprising: a printing head comprising a nozzle through which the extrudate is printed; a camera mounted on the printing head and arranged to provide visual images of the printed extrudate, and wherein the processing module is arranged to: obtain at least one visual image from the camera; perform instance segmentation on the visual image to localise and segment the extrudate; and perform distance measurement with reference to a calibration standard to determine the extrusion width of the identified printed extrudate.

16. The 3D printer of any of Claims 13 to 15, wherein the 3D printer is a multi-modal 3D printer arranged to be switchable between a fused filament fabrication - FFF- 3D printing mode and at least one electrohydrodynamic - eHD - 3D printing mode, the printer comprising: a printer frame arranged to support other components of the 3D printer; a printing head comprising at least one nozzle with electrical grounding; an insulating printing bed apparatus comprising: a conductive plate arranged to be electrified when the printer is operated in the eHD 3D printing mode; an electrically-insulated layer; an electrical insulation frame surrounding the edges of the conductive plate and arranged to isolate the conductive plate from the printer frame and to mount the conductive plate on and parallel to the electrically-insulated layer, between the electrically -insulated layer and the printing head, a thickness of the electrical insulation frame spacing the conductive plate from the electrically- insulated layer; and a spacer mounted on the opposite side of the electrically-insulated layer from the conductive plate, and comprising at least one bed-levelling adjustor; and a motion carriage arranged to move the insulating printing bed apparatus, wherein the spacer is arranged to distance the conductive plate from the motion carriage, and the electrical insulation frame is arranged to isolate the conductive plate from the motion carriage.

17. A method of automatically measuring extrusion width of an extrudate printed by a 3D printer, the method comprising: obtaining at least one visual image of the extrudate whilst a printing process is ongoing; performing machine learning-based instance segmentation on the visual image to localise and segment the extrudate; and performing distance measurement with reference to a calibration standard to determine the extrusion width of the identified printed extrudate.

18. The method of Claim 17, wherein the extrusion width is determined in real-time, and more specifically within 0.01 to 0.21 seconds of the extrudate being printed.

19. The method of Claim 17 or Claim 18, wherein the method comprises a calibration step with a calibration standard of known dimensions, the calibration step comprising, prior to obtaining the visual image of the extrudate, obtaining a visual image of the calibration standard in the same plane as the extrudate to be measured will lie, and computing an affine transformation to map the image of the calibration standard to a frame with the known dimensions of the calibration standard.

20. The method of any of Claims 17 to 19, further comprising comparing the determined extrusion width to an expected extrusion width, and stopping or adjusting printing of the extrudate if a difference between the two exceeds a set tolerance.

21. A 3D printer with integrated extrudate width measuring capabilities, the 3D printer comprising: a printing head comprising a nozzle through which the extrudate is printed; a camera mounted on the printing head and arranged to provide visual images of the printed extrudate; and a processing module arranged to: obtain at least one visual image from the camera; perform instance segmentation on the visual image to localise and segment the extrudate; and perform distance measurement with reference to a calibration standard to determine the extrusion width of the identified printed extrudate.

22. A multi-modal 3D printer arranged to be switchable between a fused filament fabrication - FFF- 3D printing mode and at least one electrohydrodynamic - eHD - 3D printing mode, the printer comprising: a printer frame arranged to support other components of the 3D printer; a printing head comprising at least one nozzle with electrical grounding; an insulating printing bed apparatus comprising: a conductive plate arranged to be electrified when the printer is operated in the eHD 3D printing mode; an electrically-insulated layer; an electrical insulation frame surrounding the edges of the conductive plate and arranged to isolate the conductive plate from the printer frame and to mount the conductive plate on and parallelto the electrically-insulated layer, between the electrically -insulated layer and the printing head, a thickness of the electrical insulation frame spacing the conductive plate from the electrically- insulated layer; and a spacer mounted on the opposite side of the electrically-insulated layer from the conductive plate, and comprising at least one bed-levelling adjustor; and a motion carriage arranged to move the insulating printing bed apparatus, wherein the spacer is arranged to distance the conductive plate from the motion carriage, and the electrical insulation frame is arranged to isolate the conductive plate from the motion carriage23. The 3D printer of Claim 22, wherein the printing head comprises at least two grounded nozzles, the at least two grounded nozzles comprising a spraying nozzle and a melt nozzle, and wherein the spraying nozzle comprises a spinneret, and the 3D printer further comprises a spinneret grounding apparatus comprising two conductive contacts biased towards each other, the biasing direction being perpendicular to the length of the spinneret, and wherein the two conductive contacts are arranged on opposing sides of the spinneret from each other such that they are biased to grip the spinneret therebetween, and each conductive contact is electrically grounded.

24. A method of retrofitting a printing bed of an FFF 3D printer to allow usage of the printing bed in at least one electrohydrodynamic - eHD - 3D printing mode, the method comprising: coating each face of the printing bed with an insulating material to form an electrically-insulated layer; mounting a conductive plate on and parallel to the electrically-insulated layer using an electrical insulation frame which surrounds the edges of the conductive plate, a thickness of the electrical insulation frame spacing the conductive plate from the electrically-insulated layer; and mounting a spacer on the opposite side of the electrically-insulated layer from the conductive plate, the spacer comprising at least one bed-levelling adjustor; and wherein the spacer is arranged to distance the conductive plate from a motion carriage of the FFF 3D printer, and the electrical insulation frame is arranged to isolate the conductive plate from the motion carriage.

25. The method of Claim 24, wherein the electrical insulation frame is made of a reversibly deformable material, and the mounting of the conductive plate on the electrically-insulated layer comprises deforming the electrical insulation frame so as to fit the conductive plate and the electrically-insulated layer therewithin, and then releasing the electrical insulation frame.

Images