Predicting green-part quality

Abstract

Aspects herein provide for predicting a physical characteristic of a potential green part, based upon a part design file, via a machine-learning model. A machine learning model is trained to output a green-part-quality vector based on a plurality of build vectors, representing values corresponding to a part design file and a set of manufacturing parameters relating to manufacturing a green part based on the part design file on a three dimensional (3D) printer. The machine-learning model receives a new build vector based on a new part file design and outputs a new green-part-quality vector representing one or more predicted physical characteristic of a potential green part, based on the new part design file.

Description

PREDICTING GREEN-PART QUALITYBACKGROUND 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 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, which provide examples as follows:

[0003] FIG. 1 illustrates an example device 100 that can facilitate implementation of systems, apparatus, methods and computer-readable storage media described herein;

[0004] FIG. 2A depicts a diagram of an example system environment for applying the computer-readable storage media suitable for implementation of aspects of the technology discussed herein;

[0005] FIG. 2B depicts a diagram of an example system environment for applying the computer-readable storage media suitable for implementation of aspects of the technology discussed herein subsequent to training of the machine-learning model;

[0006] FIG. 3 depicts a flowchart of an example method suitable for implementation of aspects of the technology discussed herein;

[0007] FIG. 4 depicts a flowchart of another example method suitable for implementation of aspects of the technology discussed herein; and

[0008] FIG. 5 depicts a flowchart of example method suitable for implementation of aspects of the technology discussed herein.DETAILED DESCRIPTION OF THE INVENTION

[0009] Additive manufacturing may be used to manufacture three-dimensional (3D) objects. 3D printing is an example of additive manufacturing. Metal-Jet Fusion (MJF) is an example of 3D printing. During an MJF 3D printing process, a 3D printed part is printed layer- by-layer in a build volume. Each layer includes powder comprised of metal spread in a desired orientation and binding agents are selectively deposited on the powder. Generally, binding agents are applied in a precise manner during the additive manufacturing process. Subsequent to application of agents, each layer may be subjected to energy from energy sources, in some aspects, such as heat lamps or laser, thereby causing at least partial removal (e.g., breakdown and elimination) of the binder. The process may be repeated for each layer until printing of the 3D part is completed. In some aspects, the 3D part is subjected to an energy source only after all the layers have been formed. Some examples of the techniques described herein may be utilized for various examples of additive manufacturing. For instance, some examples may be utilized for metal printing. Some metal printing techniques may be powder-based and driven by powder gluing and / or sintering. Some examples of the approaches described herein may be applied to area-based powder bed metal printing, such as binder jet, MJF, and / or metal binding printing, etc.

[0010] Generally, physical characteristics of parts manufactured using the additive manufacturing processes, such as the examples provided regarding metal printing, is impacted by the porosity of green parts. A green part refers to an object comprised of build material that is held together, at least, by a binding agent for example, and is formed by additive manufacturing. However, accurately measuring and predicting porosity, among othercharacteristics, has eluded some others such that defects and / or inconsistencies are not prevented; rather defects are sometimes addressed through manual trial-and-error approaches. Hereinafter, aspects of the technology provide a technological solution to predict a set of physical characteristics of a green part which may be printed according to a part design file, by utilizing a set of manufacturing parameters for manufacturing the part from the part design file. Thus, the physical characteristics of a green part to be printed from a particular part design file can be accurately predicted, which enables prediction or analysis of subsequent stages in processing the corresponding part without wasteful trail-and-error processes. Therefore, a user or manufacturer can avoid defects and / or inconsistencies in the green part and any subsequent state parts.

[0011] Some examples of the techniques described herein may utilize a machinelearning 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. In some examples of the techniques described herein, a deep neural network may predict or infer a set of physical characteristics for a potential green part anticipated to be manufactured, using additive manufacturing, based on manufacturing parameters of a new potential green part design file. The set of physical characteristics inferred for a potential green part to be generated from a green part design file may be represented by encoding and embedding the predictive data, output by a trained model, as a prediction vector, known as the green-part-quality vector. The green-part-quality vector may include indications of inferred physical characteristics of a potential green part to be generated before the potential green part is actually generated using additive manufacturing.

[0012] 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 such as memory 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 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. The machine-learning model 106 may be a neural network in some instances. For example, the machine-learning model 106 may be operative when the processor 102 executes instructions stored in the memory 104 for said model, where the machine-learning model 106, as further discussed hereinafter, may be trained and specially configured to predict green part quality. 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 to, by way of using the machine-learning model 106, predict a set of physical characteristics for a potential green part anticipated to be manufactured. The processor 102 may read and / or executing the instruction set stored on the machine-readable storage.

[0013] The device 100 may operate as part of and / or within the example system environment 200 shown in the diagram of FIG. 2A. As such, the system environment 200 may include the machine-learning model 106 of FIG. 1. The device 100 and system environment200 may be discussed herein together for clarity.

[0014] In additive manufacturing, as described herein, there are a set of manufacturing parameters 201, which reflect certain characteristics of how a part from the part design file 250 will be produced. The manufacturing parameters 201 generally comprise designed parameters 203 and observed parameters 205. The “part design file” may be referred to interchangeably with the “green part to be printed” that is defined by the part design file, for simplicity. The part design file 250 is configured to be manufactured using additive manufacturing techniques into a green part which may be subsequently processed into a final or completed part. For example, the part design file defines specifications and / or includes instructions for printing a specific part at the green state. Producing the green part is an intermediate process within an overall manufacturing process of a part described by the part design file 250. The manufacturing parameters may include and / or specify information that is determined by design, calculation, and / or measurement. Further, the additive manufacturing may be referred to as three-dimensional (3D) printing or simply as “printing.”

[0015] An encoder 212 encodes and embeds the manufacturing parameters 201 and the part design file 250 into a vector, referred to herein as a build vector at 214, as discussed below in detail. A single build vector may represent a plurality of manufacturing parameters, in some aspects. As used herein, the term “single” refers to the vector, and not necessarily to a single part to be formed from such a vector. Thus, one vector could encode manufacturing parameters and part design files for a plurality of parts, or for a singular part, depending on the manufacturing parameters 201 selected. The encoder 212 may generate a plurality of build vectors of a plurality of part designs, in various aspects.

[0016] The encoder 212 may be a neural network the operates when the processor 102 executes instructions stored in the memory 104, where the encoder 212 is trained and specially configured to utilize and transform inputs into a representation in vector form, for example, by compressing the inputs. In some examples, the encoder 212 may be a pre-trained autoencoder.The encoder 212 may utilize at least two inputs of the set of parameters of the designed parameters 203 in addition to the part design file 250, to generate the build vector 214. The designed parameters 203 are a set of print control parameters 202 and a set of build bed design characteristics 204. The first parameter comprises the set of print control parameters 202, which may be settings and / or instructions to be utilized by an additive manufacturing machine, such as a 3D printer. Some examples of print control parameters include information such as layer thickness, contone ratio, print speed, nozzle temperatures, and / or use of a cooling fan. Additional examples of print control parameters include thermal control, use of a lamp, roller speed, and / or other typical 3D printer settings. These parameters may be input from the part design file 250, may be part of the memory 104 of the device 100, may be input by a user when requesting to use the system 200, and / or may be determined by a separate 3D printer controller.

[0017] The second parameter utilized by the encoder 212 to generate a build vector 214 comprises build bed design characteristics 204. The build bed design characteristics 204 correspond to characteristics of a build platform, a build area, or a build “bed” of a 3D printer, not to be confused with a powder bed, for example. The build bed may be a solid piece of material or may have a plurality of holes or cutouts to allow for material extraction. The build bed may vary in geometry between 3D printer models and even vary between individual 3D printers as build beds may be customized to the unique outcomes desired by a particular user and / or for a particular manufacturing process. Examples of characteristics include information identifying a 3D printer, a build bed geometry, a primary orientation and / or secondary orientation of a green part, a green part design length, a green part design width, and / or a green part design height. Additional examples include potential green part design location on the build bed, a distance of potential green part design to other part(s) during a print process, and / or a part zone powder density in a location of the build bed where the green part is to be printed.As such, in examples, the encoder 212 utilizes at least one print control parameter 202, at leastone build bed design characteristic 204, and the part design file 250, to generate the build vector 214. The encoder 212 may generate the build vector 214 from additional print control parameters 201 or build bed design characteristics 204, which may improve the accuracy of the machine learning model 106.

[0018] In addition to the designed parameters 203 discussed above, observed manufacturing parameters 205 may be encoded into the build vector 214 in order to increase the accuracy of the machine learning model 106. The first of the observed manufacturing parameters 205 is a latent thermal heat vector 206. The latent thermal heat vector 206 may be encoded by a separate prediction model (not shown), as measured latent heat data may not be available until during the printing process for the part design file 250 or after the green part has been manufactured. At least a portion of a part may be printed, after which, latent thermal heat measurements may be taken to be used in subsequent modeling, known as a historical thermal value. The thermal heat measurements may be obtained by a sensor, such as a thermal sensor, thermal camera, and / or a thermopile sensor, which detect differences in temperature on a surface of the part to generate a representation of the thermal energy of a green part after printing is complete and / or during printing. A sensor may incorporate a laser to determine thermal value for discrete locations on the printed green part or portion of green part. The sensor may employ laser technology and thermopile technology to develop a two dimensional (2D) image of the printed green part, to assign temperature values to the 2D image, and to develop a heat map from the assigned temperature values of the 2D image. The “heat map” may also be referred to as a 3D thermal matrix, as the 2D image gains a third dimension, where the third dimension is a thermal value. The 3D thermal matrix may then be encoded by another device (not shown) or by the encoder 212 into a vector usable by the encoder 212 in generating the built vector 214.

[0019] Observed manufacturing parameters 205, may also include vectors, such as a solvent extraction signature vector 208. A solvent extraction signature vector 208 is a vector that encodes and / or represents at least one of a number of curing or solvent variables. Examples of curing or solvent variables include a curing profile of a particular solvent anticipated to be used in the printing process, a packing density of the build material to be printed, a build height of the part design file 250, a total solvent extraction volume, and / or a volume of the part design file 250. Additional examples of curing or solvent variables are fan usage, a print temperature, a temperature profile of a green part throughout the curing process, and a time until cured. The curing or solvent variables may be provided to a curing signature prediction network, for example, which use the variables to predict a curing station solvent extraction signature and encode it into vector form, known as the solvent extraction signature vector 208. In some examples, one or more of the curing or solvent variables may be obtained by physically measuring a particular characteristic of a partially or fully completed green part, such that the green part being measured provides the value to be used. The measured curing or solvent variable may be stored in memory and encoded in the build vector 214 for predicting the green- part-quality vector 220 of the same or similar parts by the machine-learning model 106. The measured curing or solvent variable may be used alone or with other variables previously discussed, by the curing signature prediction network to generate the curing station solvent extraction signature. The measured curing or solvent variables may thus be used directly or indirectly by encoder 212 and / or by an external encoder to produce the solvent extraction signature vector 208.

[0020] The observed manufacturing parameters 205 may comprise print powder characteristics 210 that represents physical and non-physical attributes of a printing build material of a 3D printer, such as powder build materials. Examples of the attributes include a lot number, a cycle number (e.g., a powder cycle quantity indicates a total quantity of cyclesfor which a powder build material has been used within a 3D printer), a flowability metric, and / or other detailed characteristics, for example, obtained from metrology measurements. The attributes of the printing build material may be stored within the memory 104. For example, attributes of the printing build material such as a value for a powder cycle number and a lot number may be maintained in the memory 104 of the device 100, which may be communicatively coupled to and able to monitor aspects of the 3D printer. The attributes of the printing build material may be input in part or whole by entry into the device 100, selected from a networked device to the device 100, stored in the memory 104, as part of the part design file 250, or any combination thereof.

[0021] Based on the part design file 250 used and the manufacturing parameters 201 that are selected and which relate to the part design file 250, the encoder 212 encodes and embeds values and / or vectors of such information into another vector, referred to as the build vector at 214. A single build vector 214 may represent a plurality of manufacturing parameters of a single green part design for an individual green part or a plurality of green parts, in various aspects. In some aspects, the encoder 212 may generate a plurality of build vectors 214 to represent manufacturing parameters of a plurality of green part designs, which may be printed concurrently.

[0022] Continuing with the system environment 200, the machine-learning model 106 may be trained by ingesting the build vector 214 that is created by the encoder 212. In some aspects, the machine-learning model 106 ingests a plurality of build vectors 214 representing, for example, a change in the quantity of manufacturing parameters 201. Generally, utilizing an increased quantity of manufacturing parameters relative to a lower quantity of manufacturing parameters can improve and / or increase the accuracy of machine-learning model 106. In some aspects, the machine-learning model 106 may ingest a build vector 214 of multiple different design part files, for different green parts to be printed on a 3D printer at the same time. Insome aspects, the machine-learning model 106 may access data from previous predictions, data from other models, and / or data from a prior part design, as stored in the memory 104, and may combine such stored data with the build vector 214. In some aspects, the machine-learning model 106 may select data regarding prediction models from other device and / or other 3D printers. In some aspects, the machine-learning model 106 may prompt a user to provide, input, or import additional manufacturing parameters, previously described above, to increase a level of accuracy of the machine-learning model 106.

[0023] The machine-learning model 106 outputs prediction vector(s), referred to as a green-part-quality vector(s) at 220. The green-part-quality vector represents a prediction for at least one physical characteristic of a potential green part, as defined by at least one new part design file. The machine-learning model 106 may be trained for a predetermined, fixed quantity of vector- ingestion and prediction cycles, for example. The machine-learning model 106 may be trained for a dynamic (not fixed) quantity of vector-ingestion and prediction cycles, in another example. In some aspects, the machine-learning model 106 is trained until the curve of a regression algorithm sufficiently forms a flattened plateau, 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 vector-ingestion and prediction cycles to produce a green-part- quality vector that meets or exceeds an accuracy threshold. The machine-learning model 106 may be trained until it has been validated for accuracy. Refer to FIG. 4 and description below for validation of the machine-learning model 106.

[0024] The physical characteristics of the green part to be predicted by the green-part- quality vector 220 may include the basic dimensions of the green part such as a length, a width, a height. The physical characteristics may include other physical properties of the green part such as a mass, a density, or the contour of a surface of the green part. Additionally, the physical characteristic may be based on an envelope, which surrounds the green part, such as anenvelope dimension (length, height, width) or an envelope density. Physical characteristics may also include values which relate to physical aspects of the green part such as an erosion distance, an erosion resistance, or a three-point bend test. The physical characteristics of the green part may include any measurable mechanical property such as tensile strength, shear resistance, compressive strength, and so on.

[0025] The green-part-quality vector 220 is a vector expression of at least one predicted physical characteristic of the green part. The machine-learning model 106 may rapidly be able to calculate the green-part-quality vector 220 from input vectors received by the machinelearning model 106. The green-part-quality vector may be a vector that represents the length, width, and height of the green part. The device 100 is able to quickly determine a predetermined set of physical characteristics from the value of the green-part-quality vector produced by the machine-learning model 106.

[0026] Turning to FIG. 2B, which depicts a diagram of the example system environment 200 subsequent to training, the machine-learning model 106. The machinelearning model 106 may be stored in the memory 104 and utilized to output a new green-part- quality vector 256 based on ingestion of a new build vector 254. The new build vector 254 may represent new manufacturing parameters 251 for a new part design file 252 or new manufacturing parameters 251 for the design file 250, utilized during training of the machinelearning model 106. The new build vector 254 may be formed form new manufacturing parameters 251 which includes at least one additional parameter than the manufacturing parameters 201 used in training the machine-learning model 106. A new build vector 254 using additional parameters than when the machine-learning model 106 was trained may be known as an “enhanced build vector.’’ The new design file 252 as used herein may be the same as the design file 250, however to be encoded into a new build vector 254 subsequent to training of the machine-learning model 106. A new build vector 254 may represent a plurality of newmanufacturing vectors 251 for a print bucket of new part design file 252, in some aspects. In response to ingesting the new build vector 254, the trained machine-learning model 106 outputs the new green-part-quality vector 256 that represents at least one predicted physical characteristic of the potential green part defines in the new part design file 252 prior to printing, as inferred by the machine-learning model 106. Using the new green-part-quality vector 256, for example, a user can optimize the new manufacturing parameters 251 or even the new part design file to achieve a new green part (not shown) with characteristics optimized for subsequent processing, without resorting to trial-and-error techniques. The new green-part- quality vector 256 may predict that a particular physical characteristic of the potential green part will exhibit undesirable and of poor quality, for example, due to the new manufacturing parameters 251.

[0027] Based on the new green-part-quality vector 256 that represents the quality of the green part to be manufactured, a user may modify the new manufacturing parameters 25 Ito be utilized. The encoder 212 may then generate a modified new build vector from the modified new modified manufacturing parameters (not shown) and the new part design file 252. The system 200 inputs the modified new build vector to the machine-learning model 106, and thus receives from the machine-learning model 106 a modified new green-part-quality vector. The modified new green-part-quality vector may indicate that using the modified manufacturing parameters is predicted to produce a potential green part that exhibits desirable and / or improved quality. As an example, a user may generate a set of new build vectors 254 based on a set of print build material characteristics to obtain a series of new green-part-quality vectors 256 from the machine- learning model 106. From the series of new green-part-quality vectors 256, it may be determined whether specific build materials are predicted to produce a potential green part with desired physical characteristic(s) and improved quality. An encoder 212 may generate a new build vector 254 based on anticipated changes, desirable modifications, predictions onchanges in quality, changes to input costs, time constraints, or any number of other factors to obtain a new green-part-quality vector 220. The machine-learning model 106, once trained, allows a user to input a new part design file and obtain an accurate prediction of a physical characteristic(s) of a green part that could be printed from the new part design file 252.

[0028] Predicting the physical characteristic(s) of the green part allows for pre-emptive planning and adjustments to be made prior to production of any green part. The dimensions of a green part may be used in other models to enable analysis and / or planning to subsequent manufacturing processes. For example, the green-part-quality vector, may be used by another model to determine the final dimensions of a sintered part. In another example, the green-part- quality vector, may be used by another model to predict a sinter quality, which may represent a physical characteristic of a sintered part.

[0029] FIG. 3 depicts a flowchart of an example method 300. In various aspects, the method 300 is performed by components of the example device 100 of FIG. 1 and / or components of the example system environment 200 of FIG. 2A in order to predict a physical characteristic of a green part design via a machine-learning model. At block 301 a part design file is selected by a user or by a device capable of running a machine-learning model. The part design file may have printing instructions for a 3D printer or, the part design file may be a model of a part with partial instructions or may lack instructions (no instructions). The part design file may be configured to be used to manufacture a green part using a 3D printer. For example, the part design file defines specifications and / or includes instructions for printing a part. The part design file may also contain manufacturing parameters. At block 302, a set of manufacturing parameters are selected by a user or by the device implementing the machinelearning model. The set of manufacturing parameters may be used to print a green part from the part design file. At block 304, a build vector is generated from the manufacturing parameters that are selected and the part design file that is selected. The build vector may begenerated in an encoder. At block 306, a machine-learning model is trained using a plurality of build vectors, for example, generated by the encoder. The training may iteratively performed and repeat the actions of blocks 301, 302 and 304, for example, while utilizing different part design files and / or manufacturing parameters as the selections. At block 308, the machinelearning model may generate a green-part-quality vector that predicts at least one physical characteristic of a potential green part that may be generated from the corresponding part design file using the manufacturing parameters that are selected. For example, the machine-learning model may predict one or more of the physical characteristics previously described such as a length, a width, a height, , a density, and / or a mass of, the potential green part that may be printed from the corresponding part design file. Additionally, the physical characteristic may be a contour of a surface of the potential green part as may be measured by the magnitude of difference in at least one direction of a set of points along a contour between the part design file and the potential green part.

[0030] FIG. 4 depicts a flowchart of another example method 400. In various aspects, the method 400 is performed by components of the example device 100 of FIG. 1 and / or components of the example system environment 200 of FIG. 2A in order to predict a physical characteristic of a green part design via a machine-learning model. Method 400 may be used as an extension of method 300 as depicted in FIG. 3, with the goal of validating the machinelearning model 106 for use.

[0031] At block 401 a part design file is selected by a user or by a device capable of running a machine- learning model. For example, the part design file defines specifications and / or includes instructions for printing a part. At block 402, a set of manufacturing parameters are selected for the part design file by a user or by a device capable of running a machinelearning model. The set of manufacturing parameters may be used to print a green part from the part design file. At block 404, a build vector is generated from the manufacturingparameters selected and the part design file selected. The build vector may be generated in an encoder. At block 406, a machine-learning model is trained using a plurality of build vectors, for example, generated by the encoder. The training may iteratively performed and repeat the actions of blocks 401, 402 and 404, for example, while utilizing different part design files and / or manufacturing parameters as the selections. At block 408, the machine-learning model may generate a green-part-quality vector that predicts at least one physical characteristic of a potential green part, and that may be generated from the corresponding part design file using the manufacturing parameters that are selected.

[0032] After the completion of the aspects of the method blocks described in FIG. 3, additional processes may be utilized to validate a trained machine-learning model. At block 410, the new part design file referenced in block 408 is printed / manufactured into a new green part. At block 412, the at least one physical characteristics predicted in block 408 is measured on the new part. The measurement(s) may also be known as the “ground truth.” The measurement may be taken by a measurement sensor connected to the 3D printer which prints the new part design file or device 100. The measurement sensor such as an optical sensor, visual sensor, ultrasonic sensor, laser sensor. In other examples, the measurement can be manually performed with a ruler, compass, scale or other tools known in the art to measure. The physical characteristic may be one or more dimensions of the green part design or the green part design envelope, wherein a green part design envelope is the maximum dimensions around the green part design when printed. At block 414, the measured values from the green part are compared to the predicted values for the physical characteristic predicted in the green- part-quality vector from block 408. For example, the measured values and the green part quality vector may be compared in vector form. In other aspects, the predicted physical characteristic values are extracted from the green-part-quality vector to enable the comparison of physical characteristic values. This comparison may be performed manually by a user or may becompleted by device 100 or another similar device connected which may or may not be networked with device 100. A method of comparison is to generate an error percentage value between a measured characteristic and a predicted characteristic. The error percentage value or level of precision is a percentage difference in one or more of the physical characteristics. The level of precision can be an aggregate of all of the characteristics or the largest error of the compared characteristics. At block, 416 it is determined whether the precision of the machinelearning model prediction, the green-part-quality vector, meets a predetermined threshold of accuracy. The accuracy may be less than 5% of the average physical characteristics. In some aspects, the accuracy may be less than 3%. In other aspects, the accuracy may be less than 1%. It is expected that the accuracy be between 1% and 3%. The accuracy can be relative to the amount manufacturing parameters being included in the model, the more manufacturing parameters included, the more accurate the green-part-quality vector should be. At block 420, if the green-part-quality vector meets the desired level of precision from block 416, then the machine-learning model is validated.

[0033] If the green-part-quality vector fails to meet the level of precision from block 416, at block 418 the machine-learning model 106 from block 406 is instructed to continue training. At block 418, the measurements of the printed new part design file may be incorporated into the machine-learning model 106. The measurements of the printed new part design file may also be stored in memory 104 of device 100 to provide a set of reference data for the machine-learning model 106. The machine-learning model 106 may be trained a plurality of times using a plurality of manufacturing parameters 201 and / or part design files 250 to generate a plurality of build vectors 214. Subsequent to retraining, a set of manufacturing parameters based on the new part design file, as used previously, and the new part design file are employed in blocks 402 through 408 to generate a revised green-part-quality vector. The revised green-part-quality vector may then be tested against the measurements of the physicalcharacteristic of the new green part from block 412, as stored in memory 104 of device 100. Method 400 continues from block 414 to compare the physical characteristic(s) of the revised green-part-quality vector to the measured physical characteristic(s).

[0034] In some aspects, blocks 414 through 420 may be used to validate the machinelearning model’s 106 precision of predicting individual physical characteristics. For example, certain physical characteristic(s) may have a lower pre-determined threshold for precision than others, such as length and width being having a higher accuracy target than height. As such, the machine-learning model 106, may be re-trained on only the failing physical characteristic(s) to reduce resource expenditure of device 100 or speed retraining. Once the failing physical characteristic(s) reach the desired level of precision, the machine-learning model 106 may reincorporate the previously validated physical characteristic(s) into a second new green-part- quality vector for validation after divided characteristic training. The second new green-part- quality vector would be validated as described previously.

[0035] In some aspects, method 400 can be used incrementally to assist in training of the machine-learning model 106, wherein a portion of a part design file is used to determine manufacturing parameters 201, which the encoder 212 generates a build vector 214, by which the machine-learning model 106 generates a green -part-quality vector 220. The portion of the part design file used to generate the green-part-quality vector 220 is printed in accordance with block 410, wherein method 400 continues. In some respects, method 400 can he scaled from validating the machine-learning model with printed single layers of portions of a part design file, to complete printed single layers of a green part or complete printed portions of a green part, to a complete green part, wherein the green part is based on the design file.

[0036] FIG. 5 depicts a flowchart of another example method 500. In various aspects, the method 500 is performed by components of the example device 100 of FIG. 1 and / or components of the example system environment 200 of FIG. 2A in order to predict quality ofphysical characteristics of a green part, a printed part design file before further processing as described when referring to the previous figures, via a machine-learning model. At block 502, the machine-learning model is trained to output a prediction vector, also known as a green- part-quality vector, based on a plurality of build vectors. Wherein the build vectors 214 are based upon a part design file and the manufacturing parameters 201 of the part design file, as previously described when referencing FIG. 2A. At block 504, a new build vector relating to the manufacture of a new part design file is ingested, subsequent to training the machinelearning model. The new build vector represents the new part design file, at least one print control characteristic and at least one build bed design characteristic, as discussed previously, relating to the manufacture of the new part design file. At block 506, a new prediction vector (new green-part-quality vector) is output by the trained machine-learning model that represents at least one predicted physical characteristic of the new part design if it were printed, as inferred by the machine-learning model based on the new build vector.

[0037] 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

CLAIMS1. A non-transitory machine-readable storage media having computerexecutable instructions embodied thereon that, when executed by a processor, cause the processor to: train a machine-learning model that outputs a green-part-quality vector based on a plurality of build vectors, the plurality of build vectors including values corresponding to a part design file and a set of manufacturing parameters relating to manufacturing a green part based on the part design file on a three dimensional (3D) printer; input a new build vector based on a new part file design to the machine -learning model subsequent to training; and output a new green-part-quality vector representing a predicted physical characteristic of a potential green part, based on the new part design file, as inferred by the machine-learning model based on the new build vector.

2. The machine-readable storage media of claim 1, wherein the new build vector comprises a set of print control parameters and a set of build bed design characteristics, wherein the build bed design characteristics comprises at least one value relating the new part design file to a print bed of the 3D printer.

3. The machine-readable storage media of claim 2, wherein the new build vector further comprises a latent thermal vector, representing a 3D thermal matrix, wherein the 3D thermal matrix is a predicted thermal data set for a two-dimensional representation of the potential green part.

4. The machine-readable storage media of claim 2, wherein the new build vector further comprises a solvent extraction signature, wherein the solvent extraction signature represents a predicted curing data set comprising a set of curing variables of at least a curing profile of a solvent.

5. The machine-readable storage media of claim 2, wherein the new build vector further comprises a powder characteristic value representing at least one of a composition, a lot identifier, a cycle number, or flowability of a powder that will be used or is being used to print the new part design file.

6. The machine-readable storage media of claim 1 , wherein the new green- part-quality vector comprises a predicted physical characteristic of the potential green part different than the green-part-quality vector during training.

7. The machine-readable storage media of claim 1, wherein the predicted physical characteristic comprises a length, a width, and a height of the potential green part.

8. The machine-readable storage media of claim 1, wherein the predicted physical characteristic comprises a predicted mass of the potential green part.

9. The machine-readable storage media of claim 1, wherein the predicted physical characteristic further comprises a predicted envelope density of the potential green part.

10. The machine-readable storage media of claim 1, wherein the predicted physical characteristic comprises a contour of the potential green part.

11. A method comprising: selecting a set of manufacturing parameters for a part design file configured to be printed three dimensionally (3D); printing a green part based on the set of selected manufacturing parameters and the part design file; measuring a set of physical characteristics of the green part; training a machine-learning model, using a build vector for manufacturing parameters and the part design file, to generate a green-part-quality vector representing a predicted set of physical characteristics of a potential green part, based on the part design file, as inferred by the machine-learning model; and generating, by the machine-learning model that is trained, a new green-part-quality vector that predicts a physical characteristic of a new potential green part, based on a new part design file.

12. The method of claim 11 , further comprising determining that training of the machine-learning model is complete when the green-part-quality vector attains a predetermined level of precision for the set of physical characteristics relative to the measured set of physical characteristics of the green part.

13. The method of claim 11, wherein the predetermined level of precision is less than 3 percent of difference from the measured set of physical characteristics of the green part.

14. The method of claim 11, further comprising: calibrating the machinelearning model using an enhanced build vector, wherein the enhanced build vector is based upon at least one additional manufacturing parameter than the machine-learning model was trained upon.

15. A machine-learning system comprising: a memory; a processor; and a machine-learning model stored in the memory that is pre-trained to output green-part-quality vectors, via the processor, for a part design file configured to be three dimensionally (3D) printed and that: obtains a build vector, wherein the build vector represents the part design file and a set of manufacturing parameters relating to manufacturing a green part based on the part design file on a 3D printer; and outputs a green-part-quality vector representing a predicted set of physical measurements of a potential green part, based on the part design file, inferred by the machine-learning model.

Images