Virtually calibrating a 3D printer

Abstract

A method is described in which a thermal signature corresponding to a warm-up stage of a 3D printer is identified; an encoding vector corresponding to the thermal signature is utilized to predict part defect for each part in a calibration bucket at each configuration setting of a plurality of configuration settings (e.g., an irradiance level of a plurality of irradiances levels) to apply to the calibration bucket; and, based on the predicting, a configuration setting of a plurality of configuration settings (e.g., an irradiance level of the plurality of irradiance levels) is recommended.

Description

VIRTUALLY CALIBRATING A 3D PRINTERBACKGROUND OF THE INVENTION

[0001] Three-dimensional (3D) solid parts may be produced from a digital model using additive manufacturing. Additive manufacturing may be used in rapid prototyping, mold generation, mold master generation, and short-run manufacturing. Additive manufacturing involves the application of successive layers of build material. This is unlike some machining processes that often remove material to create the final part. In some additive manufacturing techniques, the build material may be cured or fused.BRIEF DESCRIPTION OF THE DRAWING

[0002] The present technology is described in detail below with reference to the attached drawing figures, wherein:

[0003] FIG. 1 is a diagram of an example device, suitable for implementing aspects of the technology;

[0004] FIG. 2 is a block diagram of an example system, suitable for implementing aspects of the technology;

[0005] FIG. 3 is diagram of an example of a 3D printer with 22 lamps spanning over 14 zones, suitable for implementing aspects of the technology;

[0006] FIG. 4 is a flow diagram showing an example method for recommending a configuration setting based on a warm-up stage of a 3D printer, in accordance with an aspect of the technology described herein;

[0007] FIG. 5 a flow diagram showing an example method for virtually calibrating a 3D printer and printing an object, in accordance with an aspect of the technology described herein; and

[0008] FIG. 6 is a flow diagram showing an example method for recommending a configuration setting based on properties corresponding to an apparatus, in accordance with an aspect of the technology described herein.DETAILED DESCRIPTION OF THE INVENTION

[0009] Additive manufacturing systems, including those commonly referred to as “3D printers,” provide a convenient way to produce three-dimensional objects. These systems may receive a definition of a three-dimensional object in the form of an object model. This objectmodel is processed to instruct the system to produce the object. This may be performed by depositing a series of layers of a build material into a working area of the system. Chemical agents, referred to as “printing agents,” may be selectively deposited onto each layer of the build material within the working area. In some examples, the printing agents may include one or more of a fusing agent and a detailing agent, among others. Energy may be applied using a radiation source, such as an infrared lamp, to fuse areas of a layer where fusing agent has been deposited. The process may be repeated for further layers to build up a final object.

[0010] Mechanical properties and final dimensions of a part are driven by the fusing process occurring during the printing stage. Visual defects may be observed in some MJF production. In particular, two primary visual defects may be observed in MJF polymer part production, elephant skin and thermal bleed. For example, if the energy source does not supply enough energy, the object may suffer from visible strips or channels, giving the appearance of wrinkled skin, which may be referred to as "elephant skin.” Conversely, if the energy source supplies an excess of energy, a defect known as "thermal bleed" may occur, in which chunks of partially-melted build material are attached to an outer surface of the object.

[0011] Calibrating the printer relies on a trial and error process to select a fusing lamp scale factor that will result in acceptable parts. For example, in a calibration diagnostic procedure, the printer runs for a constant set of parts (a “calibration bucket”) over different fusing lamp scale factors (e.g., seven fusing lamp scale factors comprising: +6%, +4%, +2%, 0%, -2%, -4%, and -6%). For clarity, the fusing lamp scale factor corresponds to the heat emitted by the fusing lamp. The scale factor expected to produce the highest quality parts is selected by visually assessing the printed parts at each fusing lamp scale factor. In one example, the level closest to 0% where all parts are acceptable is selected. Other configuration settings rely on a similar trial and error process (e.g., selecting a temperature of the build bed).

[0012] In some examples, a 3D printer, such as a multi-jet fusion (MJF) printer includes an internal thermo-camera that captures temperature gradient at the upper layer. However, this content is private due to client data privacy policies. In contrast, printer warm-up thermal images do not reveal the parts being printed (and thus are not sensitive) and may be available for internal research and development and customer printers. Different phases in the warm-up stage provide a unique thermal signature for each printer. Some examples of the techniques described herein may utilize the warm-up signature of a printer to determine appropriate configuration settings (e.g., the appropriate fusing level irradiance or the appropriate temperature of the build bed) for a particular print job, without requiring printing a calibrationbucket. In other words, examples of the techniques described herein enable virtual calibration (e.g., 7-level calibration).

[0013] Some examples of the techniques described herein may utilize a machine learning model or models. For example, techniques described herein may utilize convolutional neural networks (CNNs) (e.g., basic CNN, deconvolutional neural network, inception module, residual neural network, etc.), recurrent neural networks (RNNs) (e.g., basic RNN, multi-layer RNN, bi-directional RNN, fused RNN, clockwork RNN, etc.), graph neural networks (GNNs), etc. Different depths of a neural network or neural networks may be utilized in accordance with some examples of the techniques described herein.

[0014] In some examples, the printer’s intrinsic thermal conditions (i.e., the “thermal signature”) are revealed during the printer warm-up. The thermal signature may comprise temperature, pulse waveform modulation (PWM), and target temperature. Time sequence vectors may be processed to obtain an encoding vector characterizing the thermal conditions of the printer. Additional signals such as ambient temperature, humidity, printer identification (e.g., serial number and firmware version), level or irradiance, build chamber wall temperature, powder temperature, melting index, and / or a part’ s center of mass may also be processed and included in the encoding vector. In some examples, the warm-up signal may be segmented and filtered according to phases.

[0015] In some examples of the techniques described herein, a deep neural network may predict or infer a configuration setting (e.g., a fusing lamp scale factor (or irradiance level) or a temperature of the build bed) that will result in acceptable parts. The fusing lamp scale factor that will result in acceptable parts may be represented by encoding and embedding the predictive data, output by a trained model, as a prediction vector. The prediction vector may include indications of inferred physical characteristics of a part (e.g., elephant skin, thermal bleed, acceptable) at various configuration settings without actually calibrating the printer or printing the part.

[0016] Beginning with FIG. 1, an example device 100 is depicted. As shown in the example of FIG. 1, the device 100 includes a processor 102, a machine-readable storage 104, and a machine learning model 106. The processor 102 may include a central processing unit (CPU), whether virtual, physical hardware, or a combination thereof. The machine-readable storage may be any electronic, magnetic, optical, or other physical storage device that stores readable and / or executable instructions. Thus machine-readable storage may be, for example,Random Access Memory (RAM), an Electrically-Erasable Programmable Read-Only Memory (EEPROM), a storage drive, an optical disc, and the like.

[0017] The machine learning model 106 may be a neural network in some instances. Examples of machine learning model types also include regression analysis, logistic regression, cluster analysis, random forest, and the like. The machine learning model 106 may be trained and deployed, for example, to perform learning that is supervised, unsupervised, reinforcement, semi-supervised, self-supervised, and the like. The device 100 includes an instruction set to cooperate with the processor 102 and the machine-readable storage (storing the machine learning model) to, by way of the machine learning model 106, virtually calibrate a printer. The processor 102 may read and / or execute the instruction set stored on the machine-readable storage. In some examples, the device 100 may be separate and distinct from the printer. In other examples, the device 100 may be integrated with and / or part of the printer.

[0018] At a high level, machine learning model 106 generally predicts, at each irradiance level, defects (or the absence of defects) in each part of a calibration bucket, based on the warm-up stage of a printer. Additionally, or alternatively, machine learning model recommends an irradiance level that will result in an acceptable part, absent elephant skin and / or thermal bleeding.

[0019] Training data may include warm-up stage vectors (e.g., temperature, PWM, target temperature, ambient temperature, humidity, printer identification, level of irradiance, build chamber wall temperature, powder temperature, melting index, and / or a part’s center of mass) and at least some of the physical characteristics (e.g., acceptable, elephant skin, and / or thermal bleed) of the part after printing, represented by another vector (the “measurement vector”). The warm-up stage vectors may be an encoded representation of various characteristics of the printer during warm-up. For example, internal and / or external sensor(s) may detect some of these characteristics (e.g., temperature, PWM, target temperature, ambient temperature, humidity, build chamber wall temperature, powder temperature, melting index, and / or a part’s center of mass). In some examples, temperature, PWM, target temperature, build wall temperature, powder temperature, humidity, and / or melting index may be encoded individually as a time series. Other characteristics (e.g., printer identification, level of irradiance) may be discrete data that can be represented by a unique embedding vector. Each of the warm up-layer vectors may be combined into an encoding vector.

[0020] The measurement vector may be an encoded representation of the physical characteristics of the part after printing. The part may be measured manually by an operatorand / or measured using various sensors (e.g., optic sensor; camera; scale) controlled by and at the direction of a computing device. An operator and / or a computing device may calculate, identify, and / or determine various physical characteristics of the parts (e.g., elephant skin, thermal bleed, or acceptable). During training, warm-up stage vectors are provided to machine learning model 106 until the output (i.e., the prediction vector(s)) aligns with the measurement vector.

[0021] The machine learning model 106 outputs prediction vector(s) that predict part defect for a bucket of parts, without initially calibrating the printer. The machine learning model 106 may be trained for a predetermined, fixed quantity of input and prediction cycles, for example. The machine learning model 106 may be trained for a dynamic (not fixed) quantity of input and prediction cycles, in another example. In some aspects, the machine learning model 106 is trained until the prediction vector(s) approaches the ground truth (i.e., the measurement vector), which is indicative of completion of the model’s learning. In yet another example, the machine learning model 106 may be trained for any quantity of input and prediction cycles to produce a prediction vector that meets or exceeds an accuracy threshold.

[0022] Subsequent to training, the machine learning model 106 may be stored in the machine-readable storagel04 and utilized to output a prediction based on input of a new warmup stage vector. For example, the stored machine learning model 106 may be utilized to output, based on input of a warm-up stage vector, a prediction of defect for a bucket of parts that may be previously unseen. The unseen bucket of parts may represent a single part or a print bucket having a plurality of parts. In response to the input of the new warm-up stage vector, the pretrained machine learning model 106 outputs a prediction vector that represents a prediction of defect for each part of the unseen bucket of parts, as inferred by the machine learning model 106. In another example, the machine learning model 106 outputs and ranks the various configuration settings based on defect probability, as inferred by the machine learning model 106 based on input of the new warm-up stage vector. Using the defect probability, in some examples, machine learning model 106 recommends a configuration setting for printing the part(s).

[0023] Machine learning model 106 may comprise a plurality of models, of the same, similar, or different types. Such a machine learning model might utilize ensemble learning, wherein a plurality of models contribute to the learning process by, for example, voting.

[0024] In some configurations, the machine learning model 106 may be embodied on servers. In other configurations, the machine learning model 106 may be implemented at leastpartially or entirely on a user device (e.g., a printer). The machine learning model 106 (and its components) may be embodied as a set of compiled computer instructions or functions, program modules, computer software services, or an arrangement of processes carried out on one or more computer systems.

[0025] The device 100 may operate as part of and / or within the example system environment 200 shown in the diagram of FIG. 2. As such, the system environment 200 may include the machine learning model 106 of FIG. 1.

[0026] In some examples, referring to FIG. 2, inputs corresponding to a printer are initially provided to calibration system 250. The inputs may include warm-up signal 202 and / or ambient conditions 204. The warm-up signal 202 may include temperature, PWM, target temperature, build chamber wall temperature, powder temperature, melting index, and / or a part’s center of mass. In some examples, sensors corresponding to a device (such as device 100 of FIG. 1) detect aspects of the warm-up signal 202. For example, sensors of the device may detect the temperature, PWM, target temperature, build chamber wall temperature, powder temperature, melting index, and / or a part’s center of mass. Ambient conditions 204 may include temperature and / or humidity. In some examples, sensors corresponding to the device detect aspects of the ambient conditions 204. In some examples, printer thermal signature encoder 206 encodes the warm-up signal 202. In some examples, ambient conditions signal encoder 208 encodes the ambient conditions.

[0027] In some examples, at least a portion of the warm-up signal and / or ambient conditions is processed as a time sequence with a one-dimensional convolutional layer and fixed width size. For example, each of temperature, PWM, target temperature, build wall temperature, powder temperature, and / or melting index may be encoded individually as a time series by printer thermal signature encoder 206 with time-sensitive features. In another example, temperature and / or humidity may be encoded by ambient conditions signal encoder 208 with time-sensitive features.

[0028] In some examples, the printer identification (e.g., serial number and firmware version) and / or the irradiance level 214 may be processed as a discrete signal into an embedding vector 216. Each discrete signal may be represented as a unique embedding vector. For clarity, the irradiance level 214 is a calibration dependent input.

[0029] In some examples, physical characteristics of the part and / or the coordinates of the part within the build bed are encoded as a vector representation of a part’s center of mass.

[0030] Each of the inputs (e.g., the encoded warm-up signal, the encoded ambient conditions, the encoded printer identification, the encoded irradiance level, and / or the encoded part’s center of mass) are combined to make up the encoding vector. The encoding vector (corresponding to the warm-up signal of the 3D printer) may be input into the trained classification network 210 to predict part defects over the various configuration settings. Based on the predicted defects, the recommender 212 determines the configuration setting (e.g., the irradiance level and / or the temperature of the build bed) that is most likely to produce acceptable parts. Accordingly, the output 218 provides the recommended configuration setting.

[0031] In some configurations, the encoder, the classification network (i.e., the classifier), and / or the recommender may be embodied on servers. In other configurations, the encoder, the classification network (i.e., the classifier), and / or the recommender may be implemented at least partially or entirely on a user device (e.g., a printer). The encoder, the classification network (i.e., the classifier), and / or the recommender may be embodied as a set of compiled computer instructions or functions, program modules, computer software services, or an arrangement of processes carried out on one or more computer systems.

[0032] In some examples, the data generated in the inference mode is used to recommend the probability of a successful completion of the next set(s) of builds. The same sequence of builds may generate different recommendations on different printers (because the thermal signature may vary from printer to printer). In some examples, a change of powder or print modes on the same printer for the same sequence of builds may generate a different recommendation.

[0033] In some examples, the recommended configuration setting (e.g., the irradiance level and / or the temperature of the build bed) may be selected and the part(s) are manufactured. For example, the objects may be manufactured by an apparatus (e.g., 3D printer) in accordance with the recommended configuration setting (e.g., the irradiance level and / or the temperature of the build bed). For instance, an apparatus may send the configuration setting (e.g., the irradiance level and / or the temperature of the build bed) to another device (e.g., 3D printer) or may execute the configuration setting (e.g., the irradiance level and / or the temperature of the build bed)to manufacture the objects based on the virtual calibration.

[0034] In some examples, the irradiance level closest to a 0% fusing lamp scale factor (i.e., the reference irradiance level) whose all parts are predicted to be acceptable is selected. It should be noted that some examples of the techniques described herein may be utilized in avariety of additive manufacturing. Some additive manufacturing techniques may be powderbased and driven by powder fusion. Some examples of the approaches described herein may be utilized in powder bed fusion-based additive manufacturing, such as Selective Laser Melting (SLM), Selective Laser Sintering (SLS), Multi-Jet Fusion (MJF), etc.

[0035] In actual experiments, a machine learning model was trained with 336 total parts in a 7-level calibration bucket (56 parts per calibration bucket). Next the machine learning model was trained with 112 parts (e.g., unseen bucket 1 and unseen bucket 2). Overall validation accuracy was 96%.

[0036] For example, for a first unseen bucket (cold), the prediction accuracy was 94%. Two irradiance levels were recommended as defect free, level 2 (+6%) and level 4 (+4%). In this experiment, level 2 agreed with the ground truth and level 4 revealed one defective part (elephant skin) in the ground truth.

[0037] In another example, for a second unseen bucket (hot), the prediction accuracy was 98%. Four irradiance levels were recommended as defect free, level 7 (-6%), level 5 (- 4%), level 3 (-2%), and level 1 (+0%). In this experiment, levels 3, 5, and 7 agreed with the ground truth and level 1 revealed one defected part (thermal bleed) in the ground truth.

[0038] Referring now to FIG. 3, an example of a 3D printer with 22 lamps spanning over 14 zones, suitable for implementing aspects of the technology, is illustrated. The energy impact to a part not only depends on the lamp in closest proximity, but also energy contributions from other lamps. For example, a part located in SW Zone 1 326 is in closest proximity to lamps 302. Accordingly, lamps 302 provide the greatest energy impact to the part. However, lamps 304, 308 may also provide an energy impact to the part. For clarity, the energy impact from the lamp(s) may or may not cause fusing of build material on which a fusing agent has been applied and may or may not result in a defect of the part.

[0039] In another example, a part located in Zone 6 330 is in closest proximity to lamp 308. Accordingly, lamp 308 provides the greatest energy impact to the part. However, lamps 302, 304, 310 may also provide an energy impact to the part.

[0040] In another example, a part located in Zone 7 328 is in closest proximity to lamp 310. Accordingly, lamp 310 provides the greatest energy impact to the part. However, lamps 302, 304, 308 may also provide an energy impact to the part.

[0041] In yet another example, a part located in NW Zone 11 322 is in closest proximity to lamps 306. Accordingly, lamps 306 provide the greatest energy impact to the part. However, lamps 304, 310 may also provide an energy impact to the part.

[0042] In each of these examples, the energy impact is determined based on the calculated energy contribution to each zone. In this way, the energy impact may be different for each part in the virtual calibration, as part of the encoded warm-up signal, based on which zone(s) the part is located.

[0043] FIG. 4 is a flow diagram showing an example method 400 for recommending a configuration setting (e.g., the irradiance level and / or the temperature of the build bed)based on a warm-up stage of a 3D printer, in accordance with an aspect of the technology described herein. The method 400 may be performed, for instance, by the example device of FIG. 1 or the example system of FIG. 2. As shown at block 402, a thermal signature corresponding to a warm-up stage of a 3D printer is initially identified. The thermal signature may be based on thermal sensing of a warm-up stage of the printer. In some examples, the warm-up stage is segmented into four phases: warm up, detected melting temperature, pressure, and printing. These phases are influential in identifying hot or cold printers that can result in two classes of defects: elephant skin (cold) and thermal bleed (hot).

[0044] In some examples, the thermal signature may comprise temperature, PWM, and target temperature. The thermal signature is encoded, at block 404, and utilized to predict part defect for each part in a calibration bucket at each configuration setting of a plurality of configuration settings to apply to the calibration bucket. In some examples, printer identification, build chamber wall temperature, powder temperature, melting index, and / or irradiance level may be encoded as part of the encoding vector. In some examples, ambient conditions comprising temperature and / or humidity and the irradiance level may be detected by sensors of the printer and encoded as part of the encoding vector. In some examples, continuous features corresponding to a center of mass of each part of the plurality of parts may be extracted and encoded as part of the encoding vector.

[0045] Based on the predicting, a configuration setting of a plurality of configuration settings, is recommended, at block 406. For example, the configuration setting may be an irradiance level and the plurality of irradiance levels corresponds to a plurality of different fusing lamp scale factors. In another example, the configuration setting may be a temperature of the build bed and the plurality of configuration settings corresponds to a plurality of different temperatures of the build bed.

[0046] FIG. 5 a flow diagram showing an example method 500 for virtually calibrating a 3D printer and printing an object, in accordance with an aspect of the technology described herein. The method 500 may be performed, in part, for instance, by the example device of FIG.1 or the example system of FIG. 2. As shown at block 502, successive layers of build material corresponding to the object to be printed in a build bed of the 3D printer are applied. In some examples, the warm-up stage of the 3D printer is initiated after the successive layers of build material are applied.

[0047] At block 504, a thermal signature corresponding to a warm-up stage of the 3D printer is detected. The thermal signature may be detected by various internal and external sensors corresponding to the printer and encoded in a vector representation. In some examples, the thermal signature comprises temperature, PWM, and target temperature during the warmup stage of the 3D printer. At block 506, part defect for the object to be printed at each configuration setting of a plurality of configurations settings is predicted. Based on the predicting, a configuration setting of a plurality of configurations settings (e.g., an irradiance level of the plurality of irradiance levels) is selected, at block 508. In some examples, the selection is automated. In other examples, the selection is made by a user. At block 510, utilizing the configuration setting, the object is printed.

[0048] FIG. 6 is a flow diagram showing an example method 600 for recommending a configuration setting based on properties corresponding to an apparatus, in accordance with an aspect of the technology described herein. The method 600 may be performed, for instance, by the example device of FIG. 1 or the example system of FIG. 2. As shown at block 602, properties corresponding to the apparatus are encoded in an encoding vector. In some examples, the properties represent an energy impact to the object at various zones during a warm-up stage of the apparatus. The properties may be encoded by a machine learning model (such as ML model 106 of FIG. 1) or an encoder (such as printer thermal signature encoder 206 or ambient conditions signal encoder 208 of FIG. 2).

[0049] A classifier utilizes the encoding vector to predict, at block 604, part defect for an object at a plurality of configuration settings. For example, part defect may be predicted by a machine learning model (such as ML model 106 of FIG. 1) or a classifier (such as classification network 210 of FIG. 2).

[0050] At block 606, a configuration setting of the plurality of configuration settings (e.g., an irradiance level or a temperature of the build bed) is recommended. For example, the configuration setting may be recommended by a machine learning model (such as ML model 106 of FIG. 1) or a recommender (such as recommender 212 of FIG. 2). In some examples, the recommender recommends the configuration setting without printing any part in the calibration bucket.

[0051] While various examples of techniques are described herein, the techniques are not limited to the examples. Variations of the examples described herein may be implemented within the scope of the disclosure. For example, operations, functions, aspects, or elements of the examples described herein may be omitted or combined.

Claims

CLAIMSWhat is claimed is:

1. Non-transitory computer storage media storing computer-readable instructions that when executed by a processor, cause the processor to perform operations, the operations comprising: identifying a thermal signature corresponding to a warm-up stage of a three- dimensional (3D) printer, the thermal signature generated by a sensor of the 3D printer; utilizing an encoding vector corresponding to the thermal signature, predicting part defect for each part in a calibration bucket at each configuration setting of a plurality of configuration settings of the 3D printer; and based on the predicting, recommending a configuration setting of the plurality of configuration settings.

2. The non-transitory computer storage media of claim 1, further comprising training a classifier to predict the part defect for each part in the calibration bucket.

3. The non-transitory computer storage media of claim 1, further comprising determining ambient conditions corresponding to the 3D printer.

4. The non-transitory computer storage media of claim 3, further comprising extracting continuous features corresponding to a center of mass of each part of the plurality of parts.

5. The non-transitory computer storage media of claim 4, further comprising encoding the thermal signature, the ambient conditions, and the continuous features in the encoding vector.

6. The non-transitory computer storage media of claim 1, further comprising encoding a serial number and firmware corresponding to the 3D printer, a build chamber wall temperature, a powder temperature, a melting index, or the plurality of irradiance levels in the encoding vector.

7. The non-transitory computer storage media of claim 1, wherein the part defect comprises one of: elephant skin, thermal bleed, or normal.

8. The non-transitory computer storage media of claim 1, wherein the plurality of configuration settings correspond to a plurality of irradiance levels or a plurality of temperatures of a build bed of the 3D printer.

9. The non-transitory computer storage media of claim 1, wherein the warm-up stage comprises a warm-up phase, a melt phase, a pressure phase, and a printing phase.

10. A method for virtually calibrating a three-dimensional (3D) printer and printing an object, the method comprising: applying successive layers of build material corresponding to the object to be printed in a build bed of the 3D printer; detecting a thermal signature corresponding to a warm-up stage of the 3D fusion printer; predicting a part defect for the object to be printed at each configuration setting of a plurality of configuration settings; based on the predicting, selecting a configuration setting of the plurality of configuration settings; and utilizing the configuration setting, printing the object.

11. The method of claim 10, wherein the thermal signature comprises temperature, pulse waveform modulation, and target temperature during the warm-up stage of the 3D printer.

12. The method of claim 10, further comprising initiating the warm-up stage of the 3D printer.

13. An apparatus, comprising:an encoder to encode, in an encoding vector, properties corresponding to the apparatus; a classifier to predict, utilizing the encoding vector, part defect for an object at a plurality of configuration settings; and a recommender to recommend a configuration setting of the plurality of configuration settings.

14. The apparatus of claim 13, wherein the recommender recommends the configuration setting without printing any part in the calibration bucket.

15. The apparatus of claim 13, wherein the properties represent an energy impact to the object at various zones during a warm-up stage of the apparatus.

Images