Predicting quality of sintered parts by means of a machine learning model

Abstract

Aspects herein provide for predicting quality of physical characteristics of a sintered part via a machine-learning model. A physical characteristic of a green part is measured and a physical characteristic of a sintered part formed from the green part is measured. A training a machine-learning model is trained using a green-state vector for the green part that is measured and a sintered-state vector of the sintered part that is measured. The machine-learning model can generate a prediction vector for a new green part, wherein the prediction vectors predicts quality of a physical characteristic of a potential sintered part to be formed from the new green part.

Description

PREDICTING QUALITY OF SINTERED PARTS BY MEANS OF A MACHINE LEARNING MODELBACKGROUND 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 DRAWINGS

[0002] The present technology is described in detail below with reference to the attached drawing figures, which provide examples as follows:

[0003] FIG. 1 depicts a diagram of an example device suitable for implementation of aspects of the technology discussed herein;

[0004] FIG. 2 depicts a diagram of an example system suitable for implementation of aspects of the technology discussed herein;

[0005] FIG. 3 depicts a flowchart of an example method suitable for implementation of aspects of the technology discussed herein;

[0006] FIG. 4 depicts a top-down view illustrating an example of erosion distance measurements between a green part and a corresponding sintered part;

[0007] FIG. 5 depicts a side view illustrating an example of erosion distance measurements between the green part and the corresponding sintered part; and

[0008] FIG. 6 depicts a flowchart of another 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. Subsequentto application of agents, each layer may optionally 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 powderbased 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, the physical quality of parts manufactured using the additive manufacturing processes, such as the examples provided regarding metal printing, is impacted by the porosity and various characteristics of green parts. Among various characteristics, porosity can result in defects and / or inconsistencies occurring in the sintered parts generated from the green parts. Accurately measuring and predicting the quality of sintered parts to be formed from green parts 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 that can utilize the physical characteristics of a green part and can predict the quality of a corresponding sintered part to be formed from the green part after sintering. Thus, by predicting the quality of a sintered part prior to sintering the green part, defects and / or inconsistencies may be predicted in a proactive manner. The predicted quality of a sintered part may utilized to determine whether a potential sintered part should be formed from a green part or scrapped, which saves sintering resources.

[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 sintered state of a sintered part that will or is anticipated to be generated from a green part. The sintered state inferred for asintered part to be generated from a green part 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 sintered part before the sintered part is actually generated from a green part.

[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 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. The machinelearning 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 to, by way of the machine-learning model 106, predict quality of physical characteristics of a sintered part via a machine-learning model. The processor 102 may read and / or execute 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. 2. As such, the system environment 200 may include the machine-learning model 106 of FIG. 1.

[0014] In some aspects, a physical characteristic of a green part is measured at 202. As used herein, a “green part” refers to an object comprised of build material that is held together, at least, by a binding agent, and is formed by additive manufacturing. Examples of a physical characteristic include a width, a length, a height, a part mass, an envelope density, an erosion distance, or any combination thereof. In aspects, a physical characteristic of a sintered part is also measured at 212, that sintered part being formed by sintering the green part that was measured prior to sintering. As used herein, a “sintered part” refers to a green part that has been sintered by the application of energy by a sintering apparatus, such as an oven or a furnace. The term “sintered part” is utilized in the context of metal additive manufacturing as anexample for the purposes of this application. The technical aspects discussed herein are equally applicable to non-metal additive manufacturing that involve densification techniques such as sintering and / or infiltration. As such, the technical aspects discussed herein apply to any process in which oven sintering is applied to a part, such as an intermediate part, which include(s) techniques of metal additive manufacturing and ceramic additive manufacturing, for example.

[0015] The green part and / or the sintered part may be measured manually by an operator and / 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 green parts and sintered parts. Additional examples of physical characteristics include volume, tensile strength, surface texture (e.g., roughness, smoothness, and particular textures), appearance (e.g., opacity, color, and gloss / sheen), conductivity, insulation properties, cross-sectional data (e.g., data obtained by scanning an interior of a part), defects, and the like. These various physical characteristics of a green part may be impactful and even determinative of the quality of a sintered part formed from that green part. Examples of defects occurring in MJF include spatter, balling phenomenon, gas porosity, surface roughness, cracking, geometric deformation such as warping, keyholes, surface-connected porosity, lack of fusion, and the like. Such defects in sintered parts may be a result of, as correlated with, and / or may be contributed to, by various physical characteristics of a corresponding green part.

[0016] Based on the measurements of the green part, an encoder 208 encodes and embeds the measurement(s) of the corresponding physical characteristic(s) into a vector, referred to as a green-state vector at 210. A single green-state vector may represent a plurality of physical characteristics of a single green part, in some aspects. The encoder 208 may generate a plurality of green-state vectors to represent a plurality of green parts, in various aspects.

[0017] Based on the measurements of the sintered part, the encoder 208 encodes and embeds the measurement(s) of the corresponding physical characteristic(s) into a vector, referred to as a sintered-state vector at 212. A single sintered-state vector may represent a plurality of physical characteristics of a single sintered part, in some aspects. The encoder 208 may generate a plurality of green-state vectors to represent a plurality of green parts, in various aspects. The vector may be generated (i.e., encoded and embedded) using a multi-layer perceptron-based module (MLP), for example.

[0018] In some aspects, variable(s) of a furnace or an oven may be measured manually by an operator and / or measured using various sensors (e.g., temperature; energy) controlled by and at the direction of a computing device, at 206. Examples of oven variables include a gas type, a gas flow rate, a sintering curve (e.g., a curve representing how, when, and how much energy is applied when sintering parts, such as timing, energy amounts, and rates for the sintering process or energy cycle), and the like. Based on the oven variables, the encoder 208 encodes and embeds the variable(s) into a vector, referred to as an oven-parameter vector at 214. Additionally or alternatively, position(s) and / or nesting locations where green part(s) may be positioned within a sintering furnace or oven may be measured and determined. The position(s) and / or nesting locations of part(s) within a sintering furnace or oven may be encoded as an oven-positioning vector and / or may be encoded within the oven-parameter vector generated from over variable(s).

[0019] In the system environment 200, the machine-learning model 106 may be trained by ingesting the green-state vector and evaluating output against the sintered- state vector, which acts as the ground truth. In some aspects, the machine-learning model 106 ingests a plurality of green-state vectors representing measurements of green parts and ingests a plurality of sintered- state vectors representing measurements of sintered parts formed from at least a portion of the green parts. In some aspects, the machine-learning model 106 may ingest the oven-parameter vector and / or an oven-positioning vector, for example, in addition to greenstate vector(s) evaluates those inputs against the ground-truth, defined by sintered-state vector(s) at the time of training.

[0020] The machine-learning model 106 outputs prediction vector(s) at 218 that predict a quality of a physical characteristic (or a plurality of physical characteristics) of a potential sintered part or parts, yet to be formed by a sintering process. 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 prediction vector(s) approach the ground truth, which is indicative of completion of the model’s learning. In yet another example, the machinelearning model 106 may be trained for any quantity of vector-ingestion and prediction cycles to produce a prediction vector that meets or exceeds an accuracy threshold. The machinelearning model 106 may be trained, for example, by ingesting green-state vectors at least until a prediction vector that is output by the machine-learning model 106 relative to sintered-statevectors of sintered parts meets or exceeds an accuracy threshold. In another example, the machine-learning model 106 may be trained by ingesting a plurality of green-state vectors at least until a prediction vector that is output by the machine-learning model 106 relative to a plurality of sintered-state vectors of sintered parts meets or exceeds an accuracy threshold. For example, the machine-learning model 106 may be determined to have completed training when the prediction vector for a part has been minimized (e.g., I predicted_value - ground_truth I measured value I). In another example where the prediction vector is determined for a print bucket, the machine-learning model 106 may be determined to have completed training when the 5-sigma value of the prediction vector for the print bucket is within a predefined threshold. In such an example the machine-learning model 106 may be determined to be trained when the 5-sigma value of the prediction vector for the print bucket deviates less than 1.3% from the targeted characteristics and / or measurements relative to the x-y axis, and less than 2.3% from the ground truth relative to the z-axis. In an example where machine-learning model 106 represents a predicted quality metric (e.g., build yield) of a predicted plurality of new sintered parts to be created from the new green parts in the print bucket, the machine-learning model 106 may include prediction algorithm(s) that classify the quality of the printing bucket based on a predetermined threshold, e.g., a binary quality classifier: good quality, bad quality; a multiclassification system: excellent quality, acceptable quality, poor quality.

[0021] Additionally or alternatively, the 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.

[0022] Subsequent to training, the machine-learning model 106 may be stored in the memory 104 and utilized to output a new prediction vector based on ingestion of a new greenstate vector. The new green-state vector may represent a single new green part or a print bucket having a plurality of new green parts. In response to the input of the new green-state vector, the pre-trained machine-learning model 106 outputs the new prediction vector that represents a predicted quality metric (e.g., dimensions; porosity) of a predicted sintered part to be created from the new green part, as inferred by the machine- learning model 106. In another example, the new prediction vector represents a predicted quality metric (e.g., build yield) of a predicted plurality of new sintered parts to be created from the new green parts in the print bucket, as inferred by the machine-learning model 106 based on the new green-state vector(s) representing the print bucket.

[0023] Using the new green-state vector, for example, a low density and high porosity new green part (e.g., value(s) for density encoded in the vector) may be predicted to produce a potential sintered part with greater defects of shrinkage and dimensional warping, and this may be output as encoded in the new prediction vector. In another example, using the new greenstate vector, a high density and low porosity new green part (e.g., value(s) encoded in the vector) may be predicted to produce a sintered part with fewer defects, such as reduced shrinkage and reduced dimensional warping, that may be output as encoded in the new prediction vector.

[0024] Based on the new prediction vector, the system environment 200 may generate and provide an indication of whether a green part or print bucket should be discarded when predicted to have a reduced ‘poor’ quality and / or yield, or should proceed to a sintering oven for the sintering process when predicted to have an increased (e.g., sufficient or ‘high’) quality and / or yield. In some examples, based on the new prediction vector, the system environment 200 may generate and provide an indication that manufacturing of a green part or print bucket should be modified (e.g., changes to parameters or materials used in the printing process) to improve the quality of the potential sintered parts that may be produced, for example. In one example, based on the new prediction vector, the system environment 200 may generate and provide an indication that post-processing of a green part or print bucket should be modified to include another densification technique (e.g., add an infiltration process to densify the build material and binder of the green part) prior to sintering to improve the quality of the potential sintered parts to be produced after sintering. In some examples, based on the new prediction vector, the system environment 200 may generate and provide an indication that sintering of a green part or print bucket should be modified (e.g., changes in the sintering process and / or oven variables / parameters such as temperature, temperature rates, sintering curve, and the like) to improve the quality of the potential sintered parts that may result from sintering. For example, the system environment 200 may generate and provide an indication that sintering of a green part or print bucket should be modified by changing the position of a part within the print bucket, changing the proximity of a part relative to another part (e.g., based on crowding; based on relative size of neighboring part), changing the proximity of a part relative to a wall of a build chamber, changing the proximity of a part relative to another part, changing the orientation (e.g., selection of a particular surface of the part can be modified by rotating the surface to face downward in the build chamber, horizontal, etc.) of a part within the sintering oven, or the like.

[0025] 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. 2 in order to predict quality of physical characteristics of a sintered part via a machine-learning model. At block 302, a physical characteristic of a green part is measured. A green-state vector is generated using the physical characteristic measured for the green part, and thus the green-state vector represents the physical characteristic measured for that green part. At block 304, a physical characteristic of a sintered part formed from the green part is measured. A sintered-state vector is generated using the physical characteristic measured for the sintered part, and thus the sintered-state vector represents the physical characteristic measured for the sintered part. At block 306, a machine-learning model is trained using a green-state vector for the green part and a sintered- state vector of the sintered part, such that the machine-learning model. At block 308, the machine-learning model 106 can generate a prediction vector that predicts quality of a potential sintered part.

[0026] For example, the machine-learning model may determine, when training, whether a length, a width, and / or a height specified in the prediction vector are within a predetermined confidence value of a length, a width, and / or a height specified in at least one sintered-state vector. The predetermined confidence value may be a numeric value that defines the closeness of the prediction vector relative to the ground truth sintered-state vector. For example, the predetermined confidence value may be specific an x-,y-, and / or axis, e.g., within a predetermined percentage of a target length value relative to the y-axis, within a predetermined percentage of a target width value relative to the x-axis, and / or within a predetermined percentage of a target height value relative to the z-axis. Additionally or alternatively, the machine-learning model may determine, when training, whether a part mass and / or a part density specified in the prediction vector are within a predetermined confidence value of a part mass and / or a part density specified in at least one sintered-state vector. Additionally or alternatively, the machine-learning model may obtain, when training, an erosion distance by quantifying a difference between a point-to-point diagonal measurement relative to a curve-to-curve diagonal measurement of an edge of a sintered part. For example, a linear measurement may be made from the interior of an edge-located comer of the part to another edge-located comer of the part, to serve as a point-to-point measurement. A linear measurement may be made from one interiorly-located corner-placed curve of the part to another interiorly-located corner-placed curve of the part as the curve-to-curve measurement,for example. The corner-placed curve corresponds to the edge-located corner subsequent to erosion occurring, i.e., where a sharp edge / corner erodes to from that comer-placed curve. In such an example, the machine-learning model may determine whether the erosion distance specified in the prediction vector is within a predetermined confidence value of an erosion distance specified in the sintered-state vector for the sintered part.

[0027] For example, FIG. 4 provides a top-down view of an example green part and a corresponding sintered part, while FIG. 5 provide a side view of the example green part and the corresponding sintered part. As shown in FIGs. 4 and 5, the example green part is shown with a lightweight outline and exhibits sharp, angled, edge-located corners designated with “x” points, while the sintered part is shown with a bold outline and exhibits soft, curved ‘corners’ due to erosion that can occur with the sintering process. In FIG. 4, the erosion distance may be calculated, for example, for each of the corresponding corners of the example green part relative to the corresponding sintered part. For example, erosion distance DI is the difference between the point-to-point diagonal measurement Pl and the curve-to-curve diagonal measurement Cl (e.g., DI = Pl - Cl); erosion distance D2 is the difference between point-to- point diagonal measurement P2 and the curve-to-curve diagonal measurement C2; erosion distance D3 is the difference between the point-to-point diagonal measurement P3 and the curve-to-curve diagonal measurement C3; and erosion distance D4 is the difference between the point-to-point diagonal measurement P4 and the curve-to-curve diagonal measurement C4. In one example, there is no erosion distance DI when Pl is equal to Cl.

[0028] Additionally or alternatively, the machine-learning model may obtain, when training, a tensile strength by performing a three point bend test that measures tensile strength based on an application of force to a midpoint of a sintered part up to failure of the part. In such an example, the machine-learning model may determine whether the tensile strength specified in the prediction vector is within a predetermined confidence value of a tensile strength specified in the sintered-state vector for the sintered part.

[0029] At block 308, a prediction vector is generated using the machine-learning model, subsequent to its training. The prediction vector is generated by the machine-learning model as representing a predicted physical characteristic of a potential sintered part to be made from the green part, in response to receipt of a new green part vector that represents a new green part for which sintering quality is desired to be predicted. In some examples, an ovenparameter vector as described herein may be generated based on a plurality of sintering oven variables and may be used by the machine-learning model during training, and / or may be usedby the trained machine-learning model to generate a new prediction vector for a new green part.

[0030] FIG. 6 depicts a flowchart of another example method 600. In various aspects, the method 600 is performed by components of the example device 100 of FIG. 1 and / or components of the example system environment 200 of FIG. 2 in order to predict quality of physical characteristics of a sintered part via a machine-learning model. At block 602, the machine-learning model is trained that outputs a prediction vector based on a plurality of greenstate and sintered-state vectors. At block 604, a new green-state vector of a new green part is ingested, subsequent to training the machine-learning model. The new green-state vector represents at least one physical characteristic of the new green part, in some aspects. At block 606, a new prediction vector is output by the trained machine-learning model that represents a predicted quality metric of a potential sintered part to be created from the new green part, as inferred by the machine-learning model based on the new green-state vector and the physical characteristic(s) encoded therein. The quality metric may be a value representing a predicted physical characteristic of the potential sintered part.

[0031] In accordance with method 300 and / or method 600, aspects herein may interpolate (e.g., linear, spline, etc.) data that may be missing or absent from a green-state vector, a sintered-state vector, an oven-parameter vector, an oven-positioning vector, and any combination thereof. Such data interpolation may be performed by components of the example device 100 of FIG. 1 and / or components of the example system environment 200 of FIG. 2 in order to predict quality of physical characteristics of a sintered part via a machine-learning model. The data interpolation techniques may be leveraged when training the machine-learning model and / or when deploying the trained machine-learning model.

[0032] In one example, the machine-learning model 106 or other components of the example device of FIG. 1 and / or system environment 200 of FIG. 2 may determine whether data is missing in a green state vector (e.g., a value(s) is missing for a physical characteristic of a green part; could not be measured or could not be accurately measured). Then, the machine-learning model 106 or other components of the example device 100 of FIG. 1 and / or system environment 200 of FIG. 2 may interpolate new data for the physical characteristic that is missing for incorporation into the green-state vector, based another green-state vector that corresponds to a neighboring green part positioned within a print bucket or a build chamber.

[0033] In one example, the machine-learning model 106 or other components of the example device of FIG. 1 and / or system environment 200 of FIG. 2 may determine whetherdata is missing in a sintered-state vector (e.g., a value(s) is missing for a physical characteristic of a sintered part; could not be measured or could not be accurately measured). Then, the machine-learning model 106 or other components of the example device 100 of FIG. 1 and / or system environment 200 of FIG. 2 may interpolate new data for the physical characteristic that is missing in the sintered-state part for incorporation into the sintered-state vector based another sintered-state vector that corresponds to a neighboring sintered part positioned within a furnace or an oven.

[0034] The data interpolation process may be repeated for various missing data and for multiple green-state and / or sintered-state vectors. An estimated value may be generated for missing data corresponding to a physical characteristic measurement by using at least one existing value for that corresponding physical characteristic that is / has been measured for a neighboring part or parts (e.g., encoded in a vector) that are positioned in a print bucket, build chamber, or furnace proximate to the part for which the data is missing.

[0035] When generating an estimated value for missing data of a physical characteristic measurement, neighboring parts may identified as value sources for calculating an estimated value based on the neighboring parts’ positioning by degree of proximity (e.g., distance from a neighboring part to the part for which the data is missing) and / or in a particular plane (e.g., x-axis, y-axis, z-axis). For example, neighboring parts that are positioned closer (e.g., shortest distance) to the part for which the data is missing may be identified as optimized sources for physical characteristics measurements to be used to calculate an estimate value for the missing data. Additionally or alternatively, neighboring parts which are positioned similarly within the print bucket or furnace (e.g., where the part for which the data is missing and neighboring parts are located adjacent to each other and to a wall of the print bucket or furnace; where the part for which the data is missing and neighboring parts are located adjacent to each other and within an interior of the print bucket or furnace) to the part for which the data is missing may be identified as optimized sources for physical characteristics measurements to be used to calculate an estimate value for the missing data, in an example. The neighboring parts may be identified as experiencing similar printing or sintering conditions based on their proximity to the part for the which the data is missing, such that the neighboring parts are predicted to have the same or similar values for the same or similar physical characteristics for which data is missing.

[0036] When generating an estimated value for missing data of a physical characteristic measurement, neighboring parts may be identified as value sources for the estimated valuebased on the neighboring parts’ directional positioning in the x, y, or z-axis as related to the physical characteristic for which the data is missing, in addition to proximity to the part for which the data is missing. For example, when the missing data corresponds to a physical characteristic associated with the z-axis (e.g., height), the directional proximity of neighboring parts relative to the z-axis may be preferentially used when determined which neighboring part(s) should be utilized to generate the estimated value. In one example, when the missing data corresponds to a physical characteristic associated with the x- or y-axis (e.g., length or width), the directional proximity in the x- and / or y-axis of neighboring parts may be used when determined which neighboring part(s) should be utilized to generate the estimated value.

[0037] The directional proximity may be considered selecting neighboring parts as value sources for estimating values because the quality of various physical characteristics of a sintered part may vary distinctly by the particular axis. For example, the quality of particular physical characteristics may be impacted (e.g., due to z-axis powder layer stacking direction; part build orientation relative to z-axis) along the z-axis due to the layer-by-layer manufacture of the part, printhead specifications, build materials, and / or binder properties. Examples of build material properties that may impact the quality of the part include composition, particle mean size, particle size distribution, spherical morphology, spreadability, flow and agglomeration, packing density, chemistry, and the like. In another example, the quality of particular physical characteristics may be impacted (e.g., due to layer spreading along the x- and / or y-axis via a recoater, blade, roller, or the like; part build orientation relative to x- and y- axis) along the x- and / or y-axis due to the manufacturing process, printhead specifications, build materials, and / or binder properties. Due to these considerations, neighboring part(s) may be identified and selectively chosen when interpolating an estimated value for missing data, so that the estimated value may be inserted into a green-state or sintered- state vector, whether when training the model or when deploying the trained model. Accordingly, missing data may be backfilled intelligently and utilized such that the machine-learning model produces accurate predictions of part quality. Historical data from other batches, green parts, and / or ovens may be used to interpolate data for missing data in over-parameter vectors and / or oven-positioning vectors.

[0038] Although green-state and sintered-state vectors are discussed as examples, the same or similar data interpolation processes can be performed to manage missing data in an oven-parameter vector and / or an oven-positioning vector. For example, oven parameter information, sintering process curves, positions (e.g., x, y, and z location; height of shelf inoven, specific oven slot) of other parts in the same batch, and / or oven parameters from other same or similar batches could be used to fill in missing data.

[0039] 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 method comprising: measuring a physical characteristic of a green part; measuring a physical characteristic of a sintered part formed from the green part; training a machine-learning model using a green-state vector for the green part and a sintered-state vector of the sintered part, the machine-learning model for generating a prediction vector that predicts quality of a physical characteristic of a potential sintered part; and generating a prediction vector using the machine-learning model subsequent to training.

2. The method of claim 1, wherein the green- state vector represents the physical characteristic measured for the green part, wherein the sintered-state vector represents the physical characteristic measured for the sintered part, and wherein the physical characteristic is a width, a length, a height, a part mass, an envelope density, an erosion distance, or any combination thereof.

3. The method of claim 1, further comprising: generating the green-state vector using the physical characteristic measured for the green part; and generating the sintered-state vector using the physical characteristic measured for the sintered part.

4. The method of claim 1, wherein the prediction vector is generated by the machine-learning model as representing a predicted physical characteristic of a predicted sintered part.

5. The method of claim 1, wherein training the machine- learning model comprises: determining whether a length, a width, and a height specified in the prediction vector are within a predetermined confidence value of a length, a width, and a height specified in the sintered-state vector.

6. The method of claim 1, wherein training the machine- learning model comprises: determining whether a part mass and a part density specified in the prediction vector are within a predetermined confidence value of a part mass and a part density specified in the sintered-state vector.

7. An apparatus comprising: a processor; a machine-readable storage coupled to the processor; and an instruction set to cooperate with the processor and the machine-readable storage to: train a machine-learning model that outputs a prediction vector based on a plurality of green-state vectors and a plurality of sintered-state vectors, the plurality of green-state vectors representing physical characteristics of green parts and the plurality of sintered-state vectors representing physical characteristics of sintered parts; ingest a new greenstate vector of a new green part subsequent to training the machine-learning model; and output a new prediction vector representing a predicted quality metric of a potential sintered part to be created from the new green part, as inferred by the machine-learning model based on ingesting the new green-state vector.

8. The apparatus of claim 7, wherein a length, a width, and a height specified in the prediction vector is within a predetermined confidence value of a length, a width, and a height specified in at least one of the sintered-state vectors.

9. The apparatus of claim 7, wherein a part mass and a part density specified in the prediction vector is within a predetermined confidence value of a part mass and a part density specified in at least one of the sintered-state vectors.

10. The apparatus of claim 7, the instruction set to cooperate with the processor and the machine-readable storage to: obtain an erosion distance by quantifying a difference between a point-to-point diagonal measurement relative to a curve-to-curve diagonal measurement of an edge of a sintered part; and determining whether the erosion distance specified in the prediction vector is within a predetermined confidence value of an erosion distance specified in the sintered-state vector for the sintered part.

11. The apparatus of claim 7, the instruction set to cooperate with the processor and the machine-readable storage to: obtain a tensile strength using a three point bend test, the three point bend measuring the tensile strength based on an application of force at a midpoint of the sintered part; and determining whether the tensile strength specified in the prediction vector is within a predetermined confidence value of a tensile strength specified in the sintered-state vector for the sintered part.

12. The apparatus of claim 7, wherein the green-state vectors include physical characteristics of green parts, and wherein the instruction set cooperates with the processor and the machine-readable storage to, for at least one of the green state vectors: determine data for at least one physical characteristic is missing in the at least one green-state vector; and interpolate new data for the at least one physical characteristic that is missing for incorporation into the at least one green-state vector based another green-state vector that corresponds to a neighboring green part in a build chamber.

13. The apparatus of claim 7, wherein the sintered- state vector includes physical characteristics of the sintered part, and wherein the instruction set cooperates with the processor and the machine -readable storage to, for the sintered-state vectors: determine data for at least one physical characteristic is missing in at least one of the sintered-state vectors; and interpolate new data for the at least one physical characteristic that is missing in at least one of the sintered-state part for incorporation into the at least one sintered-state vector based another sintered-state vector that corresponds to a neighboring sintered part.

14. The apparatus of claim 7, the instruction set to cooperate with the processor and the machine-readable storage to: instructions to generate an oven-parameter vector based on a plurality of sintering oven variables, wherein the new prediction vector is generated further based on the oven-parameter vector.

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 prediction vectors based on ingestion of a plurality of green-state vectors representing a plurality of physical measurements of green parts and a plurality of sintered-state vectors representing a plurality of physical measurements of sintered parts, the machine-learning model to, via the processor, for a new individual green part or a print bucket having a plurality of new green parts: ingest a new green-state vector; and output a new prediction vector representing a predicted quality metric of a potential sintered part as inferred by the machine-learning model; and generating and providing, based on the new prediction vector, an indication to discard the new green part.

Images