阅读说明：本技术 基于热模型和传感器数据在增材制造过程中校正构建参数的系统和方法 (System and method for correcting build parameters in an additive manufacturing process based on thermal models and sensor data ) 是由 苏布拉吉特·罗伊乔杜里 亚历山大·陈 平小虎 贾斯汀·小甘伯恩 托马斯·西崔尼蒂 布莱恩· 于 2020-01-15 设计创作，主要内容包括：向增材制造机器提供更新的构建参数,以提高由机器制造的零件的质量。在使用第一组构建参数制造零件期间,从增材制造机器接收传感器数据。接收第一组构建参数。基于第一组构建参数和接收到的传感器数据确定评估参数。基于从第一组构建参数导出的零件的热模型生成热数据。将第一算法应用于接收到的传感器数据、确定的评估参数和生成的热数据以产生第二组构建参数,第一算法被训练以改进评估参数。将第二组构建参数输出到增材制造机器以生产第二零件。(Updated build parameters are provided to an additive manufacturing machine to improve the quality of parts manufactured by the machine. During manufacturing of a part using a first set of build parameters, sensor data is received from an additive manufacturing machine. A first set of build parameters is received. An evaluation parameter is determined based on the first set of build parameters and the received sensor data. Thermal data is generated based on a thermal model of the part derived from the first set of build parameters. A first algorithm is applied to the received sensor data, the determined evaluation parameters, and the generated thermal data to produce a second set of build parameters, the first algorithm being trained to improve the evaluation parameters. Outputting the second set of build parameters to an additive manufacturing machine to produce a second part.)

1. A method for providing updated build parameters to an additive manufacturing machine, the method comprising:

receiving sensor data from the additive manufacturing machine via a communication interface of a device comprising a processor during manufacturing of a part using a first set of build parameters;

receiving the first set of build parameters;

determining, using the processor of the device, an evaluation parameter based on the first set of build parameters and the received sensor data;

generating, using the processor of the apparatus, thermal data based on a thermal model of the part derived from the first set of build parameters;

applying, using the processor of the device, a first algorithm to the received sensor data, the determined evaluation parameters, and the generated thermal data to produce a second set of build parameters, the first algorithm being trained to improve the evaluation parameters; and

outputting the second set of build parameters to the additive manufacturing machine to produce a second part.

2. The method of claim 1, wherein the evaluation parameter comprises a quality score determined by applying a second algorithm to the first set of build parameters and the received sensor data.

3. The method of claim 2, wherein the second algorithm is trained by receiving a reference derived from physical measurements made on at least one reference part, the at least one reference part being constructed using a set of reference construction parameters.

4. The method of claim 1, wherein the generating of the thermal data comprises calculating a first set of thermal data values based on a nominal thermal model and the first set of build parameters.

5. The method of claim 4, wherein the generating of the thermal data further comprises:

determining an updated thermal model based on a comparison of the calculated first set of thermal data values and the received sensor data; and

a second set of thermal data values is calculated based on the updated thermal model.

6. The method of claim 4, wherein the nominal thermal model is derived by:

dividing the volume of the part into voxels;

determining the relative amount of surrounding material within a defined radius of the center of each of the voxels; and

calculating a thermal data value for each voxel based on the relative amount of the surrounding material.

7. The method of claim 1, wherein the sensor data is received from at least one of a laser power sensor, an actuator sensor, a melt pool sensor, and an environmental sensor.

8. A system for providing updated build parameters to an additive manufacturing machine, the system comprising:

an apparatus comprising a communication interface, the apparatus configured to receive sensor data from the additive manufacturing machine during manufacturing of a part using a first set of build parameters, the apparatus further comprising a processor configured to:

receiving the first set of build parameters;

determining an evaluation parameter based on the first set of build parameters and the received sensor data;

generating thermal data based on a thermal model of the part derived from the first set of build parameters;

applying a first algorithm to the received sensor data, the determined evaluation parameters, and the generated thermal data to produce a second set of build parameters, the first algorithm being trained to improve the evaluation parameters; and

outputting the second set of build parameters to the additive manufacturing machine to produce a second part.

9. The system of claim 8, wherein the evaluation parameter comprises a quality score determined by applying a second algorithm to the first set of build parameters and the received sensor data.

10. The system of claim 9, wherein the second algorithm is trained by receiving a reference derived from physical measurements made on at least one reference part, the at least one reference part being constructed using a set of reference construction parameters.

11. The system of claim 8, wherein the generation of the thermal data comprises calculating a first set of thermal data values based on a nominal thermal model and the first set of build parameters.

12. The system of claim 11, wherein the generation of the thermal data further comprises:

determining an updated thermal model based on a comparison of the calculated first set of thermal data values and the received sensor data; and

a second set of thermal data values is calculated based on the updated thermal model.

13. The system of claim 11, wherein the nominal thermal model is derived by:

dividing the volume of the part into voxels;

determining the relative amount of surrounding material within a defined radius of the center of each of the voxels; and

calculating a thermal data value for each voxel based on the relative amount of the surrounding material.

14. A non-transitory computer readable storage medium storing program instructions that, when executed, cause a processor to perform a method for providing updated build parameters to an additive manufacturing machine, the method comprising:

receiving sensor data from the additive manufacturing machine via a communication interface of a device including the processor during manufacturing of a part using a first set of build parameters;

receiving the first set of build parameters;

determining, using the processor of the device, an evaluation parameter based on the first set of build parameters and the received sensor data;

generating, using the processor of the apparatus, thermal data based on a thermal model of the part derived from the first set of build parameters;

applying, using the processor of the device, a first algorithm to the received sensor data, the determined evaluation parameters, and the generated thermal data to produce a second set of build parameters, the first algorithm being trained to improve the evaluation parameters; and

outputting the second set of build parameters to the additive manufacturing machine to produce a second part.

15. The computer-readable storage medium of claim 14, wherein the evaluation parameter comprises a quality score determined by applying a second algorithm to the first set of build parameters and the received sensor data.

Technical Field

The disclosed embodiments are directed to correcting build parameters in an additive manufacturing process based on a thermal model and sensor data.

Background

The term "additive manufacturing" refers to a process for synthesizing a three-dimensional object, wherein successive layers of material are formed by an Additive Manufacturing Machine (AMM) under computer control to create the object using digital model data from a 3D model. One example of additive manufacturing based on powder bed fusion is Direct Metal Laser Sintering (DMLS), which uses a laser injected into a powder metal bed that is automatically aimed at a point in space defined by a 3D model, thereby melting the materials together to create a solid structure. The term "direct metal laser melting" (DMLM) may more accurately reflect the nature of the process, as it typically achieves a fully mature homogeneous melt pool and a fully dense mass upon solidification. The rapid localized heating and cooling properties of the molten material enable near wrought material properties to be achieved after application of any necessary heat treatment.

The DMLM process uses a 3D Computer Aided Design (CAD) model of an object to be manufactured, whereby a CAD model data file is created and sent to a manufacturing facility. The technician can use the 3D model to correctly position the geometry for the part build and can add support structures to the design as needed. Once this "build file" is complete, it is "sliced" into layers of appropriate thickness for the particular DMLM manufacturing machine and downloaded to the machine to allow the build to begin. DMLM machines use, for example, 400W Yb fiber lasers. Within the build chamber area, there is a powder dispensing platform and a build platform, and a recoater blade for moving new powder on the build platform. The metal powder is fused into a solid part by locally melting the metal powder using a focused laser beam. In this way, the part is built up layer by layer in superposition — typically using layers of 20 to 100 microns thick. This process allows the automated creation of highly complex geometries directly from 3D CAD data without any tools. DMLM produces parts with high precision and detail resolution, good surface quality and excellent mechanical properties.

In a DMLM process, anomalies such as subsurface porosity, cracks, lack of fusion, etc. can occur due to various machine, programming, environmental, and process parameters, as well as due to the chemistry of the materials used. For example, imperfections in the machine calibration of the mirror position and laser focus may result in a heavily filled (bulk-fill) laser channel not intersecting the edge profile channel. Such defects may result in unfused powder near the surface of the component, which may break through the surface resulting in anomalies that cannot be repaired by post-treatment heat treatment steps including Hot Isostatic Pressing (HIP). Laser and optics degradation, filtering, and other typical laser welding effects can also significantly affect process quality, especially at tens or hundreds of hours per build operation.

In conventional additive manufacturing practices, Part Build Plans (PBPs) are generated for a particular part design and executed by an Additive Manufacturing Machine (AMM). Based on the PBP, the AMM controls a number of build parameters that are applied during the build, including the travel path of the material addition zone and parameters that govern the application and processing of material added to parts in the zone. Often, there is a complex relationship between these parameters and the quality of the build part.

The design of the PBP is an iterative process that includes building a part based on a test PBP, then evaluating the resulting test part quality, and then modifying the test PBP to adjust for the desired part quality. Iterations of such a test PBP to meet overall manufacturing requirements (e.g., part quality and productivity) may require multiple iterations to achieve the desired manufacturing requirements. Evaluation of the quality of a test part is conventionally accomplished by experimentally testing the part using destructive or non-destructive techniques. In particular, a DMLM part may be sectioned, an optical micrograph generated from the processed section, and the micrograph processed to quantify the anomaly. The evaluation of the quality of the test parts is based on such tests. Such testing is laborious, expensive, and time consuming, and significantly increases the time and cost of developing an acceptable PBP for distribution to the final product.

In conventional methods, a part is built using a fixed set of parameters, and then various physical measurements, such as cutting/microscopic analysis, Coherence Tomography (CT) scanning, and other inspection techniques, are performed to assess the quality of different regions of the part. A subsequent construction then takes place, wherein the geometry can be segmented and assigned different sets of parameters. The built part is physically tested and further iterations are performed until the part quality converges to within an acceptable range. Each such iteration may take, for example, 3-4 weeks, as the part may need to be sent to a specialized facility for qualification, i.e., physical measurement, and then a design expert must interpret the qualification results and make decisions regarding segmentation and parameter changes. Such methods may require 10 to 12 iterations, which means that a year or more may be required to produce an acceptable part. Another disadvantage of the manual segmentation/parameter correction method is that the boundaries of the power stage tend to be possible points of failure.

Disclosure of Invention

The disclosed embodiments provide a method for correcting predictable disturbances in a DMLM process using a combination of models and sensor data. Based on the time scale of the interference prediction, the technique can be used to improve part quality from build to build (e.g., for geometric interference) or even layer to layer (e.g., for smoke occlusion). The goal is to reduce the cycle time for part parameter optimization, which is conventionally done by trial and error, so several weeks may be required to converge to an acceptable set of parameters. Furthermore, conventional methods are more artistic than scientific, as the end result depends on the expertise of the individual.

In the disclosed embodiments, an initial guess of the scan parameters is estimated based on a model that can be performed quickly. The results of the iteration are recorded in the sensor and compared to a previously generated reference (e.g., the results of a previous iteration or the output of the model). The estimation error is then fed back through a tracking filter to improve the model, and the updated model is used to generate a new set of scan parameters. The tracking error can then be used to further tune the scan parameters as needed. Due to the algorithmic approach, it is expected that the process converges to the optimal set of parameters after only a few iterations in a single construction. Furthermore, in conventional methods, the iteration and adjustment are manual, and thus the results depend on the expertise of the engineer. The algorithmic approach provides better results without direct human intervention. In some conventional methods, the iteration time per cycle is several weeks, as the iteration results are evaluated by post-construction cutting and material identification. In contrast, using the algorithms described herein, results can be evaluated from sensor data immediately after construction. Since the parameters are adjusted algorithmically, rather than by trial and error, fewer iterations are required for convergence. Conventional manual parameter optimization requires borrowing materials because they cannot segment parts with sufficiently fine resolution to have different scan parameters along the run (strike).

Lower New Product Introduction (NPI) cycle times result in cost savings for complex parts and greater throughput (i.e., more parts are optimized at the same time). The techniques described herein also have the potential to expand design space by implementing geometries that were not otherwise possible. Combining sensor data with model data and updating the model using a tracking filter may yield higher fidelity results than conventional approaches.

The disclosed embodiments provide for predicting part quality without the need for physical testing steps in each trial build iteration. Part quality models are developed based on sensor measurements made during part build and other information known at build time. Part quality-based decisions, such as modifications to the PBP, or part acceptance/rejection, are based on quality model results. Analyzing data generated during build and known at the time of analysis, rather than performing post-build testing, reduces the cost and elapsed time of PBP development, as well as the cost and time for quality evaluation of production parts. The disclosed methods may replace physical testing in whole or in part, or may replace certain portions of the entire testing process (e.g., long build times and expensive portions). These methods may be applied to selected iterations, and physical testing is used in the selection of iterations. These methods can be used to screen build parts and reduce the number of parts that are subjected to physical testing. In disclosed embodiments, the output of the mass fraction generator may be appended to the entire part, a section of the part, the entire build, or a section of the build. The output of the mass fraction generator may be appended to portions of the part, such as to contours, thin walls, and salient portions, rather than to bulk regions, based on complexity or geometry. The output of the quality score generator may be binary, such as pass/fail, or may have multiple levels, such as high, medium, and low, in which case parts with high quality scores may be considered good parts, parts with medium quality scores may be considered acceptable parts (i.e., parts for use in less critical applications), and low quality parts may be rejected. The output of the mass fraction generator may be a set of values indicating the particular type of post-processing (e.g., post-processing a, post-processing B, and rejection) required for the part. For example, some parts may require mild Hot Isostatic Pressing (HIP) processing, other parts may require aggressive HIP, other parts may cut away useful sections, and other parts may be rejected.

In the disclosed embodiments, the photodiode response to anomalies and microstructural changes in the finished condition caused by process variables may be considered, including the correlation between mass fraction and microstructure of the post-processed part (e.g., a part subjected to heat treatment and/or HIP). The material microstructure and chemistry of an additive manufactured part or representative region/section of a part may be measured/mapped by direct (e.g., optical, SEM imaging) or indirect methods (e.g., diffraction, spectroscopy), where the output may be a single value or a set of single values for each measurement type (e.g., mean, median, standard deviation, etc.) or a full field spatial profile of the measurement region. Methods for mapping/measuring the microstructure and elemental chemical distribution of finished and post-processed additively manufactured parts may include, for example, optical imaging, scanning electron imaging, backscattered electron imaging, electron backscatter diffraction, energy or wavelength dispersive spectroscopy, atomic force microscopy, X-ray diffraction, transmission electron microscopy (both imaging and diffraction), and the like.

In the disclosed embodiments, the mass fraction may be determined for different anomalies, such as hole density, crack density, and unfused defect density. The individual total scores may be derived from a combination of multiple sub-scores, such as a sum, a weighted sum, a maximum, an average, a weighted average. A response map formed of multiple surface images may be generated in which the mass fraction is mapped to input process parameters (e.g., laser power, scan speed, beam spot size/focus offset, and hatch spacing). A response map may be generated in which finishing anomalies (e.g., pores, cracks, unfused defects) and microstructure measurement parameters (e.g., grain size and gamma/gamma prime size distribution) are mapped as a function of input parameters (e.g., laser power, scan speed, beam spot size/focus offset, hatch spacing), derived values (e.g., linear heat input, energy density, and beam intensity), process variables (e.g., molten pool width and depth), and generated mass fractions.

In one aspect, the disclosed embodiments provide a method (and corresponding system and software) for providing updated build parameters to an additive manufacturing machine. The method includes receiving sensor data from an additive manufacturing machine via a communication interface of an apparatus including a processor during manufacturing of a part using a first set of build parameters. The method further includes receiving a first set of build parameters. The method further includes determining, using a processor of the device, an evaluation parameter based on the first set of build parameters and the received sensor data. The method further includes generating, using a processor of the apparatus, thermal data based on a thermal model of the part derived from the first set of build parameters. The method further includes applying, using a processor of the device, a first algorithm to the received sensor data, the determined evaluation parameters, and the generated thermal data to produce a second set of build parameters, the first algorithm being trained to improve the evaluation parameters. The method further includes outputting the second set of build parameters to an additive manufacturing machine to produce a second part.

Embodiments may include one or more of the following features.

The evaluation parameter may include a quality score determined by applying a second algorithm to the first set of build parameters and the received sensor data. The second algorithm may be trained by receiving a reference derived from physical measurements made on the at least one reference part, the at least one reference part being constructed using a set of reference construction parameters. The generation of thermal data may include calculating a first set of thermal data values based on the nominal thermal model and the first set of build parameters. The generation of thermal data may include determining an updated thermal model based on a comparison of the calculated first set of thermal data values with the received sensor data; and calculating a second set of thermal data values based on the updated thermal model. The nominal thermal model may be derived by: dividing the volume of the part into voxels; determining a relative amount of surrounding material within a defined radius of the center of each voxel; and calculating a thermal data value for each voxel based on the relative amount of surrounding material. Sensor data may be received from at least one of a laser power sensor, an actuator sensor, a weld pool sensor, and an environmental sensor.

Drawings

FIG. 1 is a block diagram of a system for correcting build parameters for producing a part in an additive manufacturing process;

FIG. 2 is a diagram of a system for determining a mass fraction of a produced part using a mass fraction generator that receives sensor data from an additive manufacturing machine;

FIG. 3 is a schematic diagram of a quality score generator;

FIG. 4 depicts a system for updating a thermal model of a part produced in an additive manufacturing process;

FIG. 5A depicts a portion of a part voxelized to produce a nominal thermal model;

FIG. 5B depicts in a top view the relative amount of surrounding material within a defined radius that determines the center of a particular voxel;

FIG. 5C depicts in a side view the relative amount of surrounding material within a defined radius that determines the center of a particular voxel; and

fig. 6 is a block diagram of a system according to some embodiments.

Detailed Description

Additive manufacturing build using a set of parameters that are fixed for all positions in the geometry of the part may not produce satisfactory results. For example, assume a design using a fixed set of parameters developed based on the characteristics of the materials used in the manufacture of the part. Such a set of parameters may work well in a large area of the part (i.e., a portion with a relatively uniform geometry). However, in thin-walled portions of the part where the thermal conductivity is much lower than in the bulk region, the weld pool size will be larger and the weld pool will be hotter, so the material properties in the thin-walled region may differ significantly from those in the bulk region, which may result in unsatisfactory part quality when using a fixed set of parameters. As a further example, if the part has a protruding surface, e.g., a dome shape, there is little thermal conductivity. Consequently, the weld pool will be relatively much larger, which results in a part being built with a very poor surface finish. Therefore, a build using a nominal set of parameters may result in imperfections in the material properties. The set of nominal parameters may be adjusted in an attempt to improve the properties of the surface of the material. For example, the laser power may be reduced throughout the entire build or in segmented areas. However, such adjustments can introduce or increase the porosity of the material.

In the disclosed embodiment, Iterative Learning Control (ILC) is used during the design phase to apply variable corrections to build parameters for predictable disturbances to correct laser power levels, for example, according to laser position. ILCs are particularly useful for part geometries where the thermal conductivity varies significantly in different parts of the geometry. With ILC, there is finer control over the build parameters, so fine-grained regions can be controlled separately. This configuration helps to minimize the introduction of porosity because of the finer control of the laser power.

Fig. 1 is a block diagram of a system 100 for correcting build parameters for producing a part in an additive manufacturing process. The system uses a thermal model and evaluation parameters (e.g., quality scores) derived from the build parameters and sensor data. A set of nominal build parameters is input to an Additive Manufacturing Machine (AMM)110, for example, a Direct Metal Laser Melting (DMLM) printer (i.e., machine). The nominal build file 120 may be in the form of a Common Layer Interface (CLI) file that may include a set of scan parameters, i.e., build parameters. When DMLM printer 110 is built based on nominal build file 120, sensor data 130 is collected from various sensors associated with the printer. The sensor data 130 is input to a quality score generator 140, and the quality score generator 140 calculates a score indicative of the quality of the build part without extensive physical testing of the part.

The nominal build file 120 is also input to a thermal model 150, the thermal model 150 modeling the thermal response of the build part to the applied laser power. As described in further detail below, thermal model 150 uses nominal build file 120 and sensor data 130 received from DMLM printer 110 to predict the thermal density within the volume of build part 155 that would result from applying a particular level of heat input from the laser during scanning. In effect, the thermal model 150 creates a correlation between the thermal input parameters (e.g., laser power and scan speed) specified in the build file at each location in the scan path and the expected sensor readings (e.g., photodiode readings) for that location during build.

The sensor data 130, the mass scores calculated by the mass score generator 140, and the outputs of the thermal model are input to an Iterative Learning Control (ILC) 160. The ILC 160 uses machine learning algorithms to generate updated build files 170 based on these inputs, as described in further detail below. Thus, the ILC 160 creates a mapping between the scan parameters of the build file and the resulting quality scores of the parts produced using the build file, which allows the build file to be optimized using an iterative machine learning process. This process results in a higher quality of the built part without the need for multiple experimental tests as is done with conventional methods.

Iterative learning control 160 is a term that encompasses various learning and control algorithms that are structured to learn from a previous build and improve the quality of a subsequent build. The disclosed embodiments provide for the use of quality scores in control applications that require reference to track through the use of an iterative learning process. As discussed herein, the generation and use of mass fractions allows for modeling of an array of physical properties, e.g., porosity, surface finish, etc., which are conventionally determined using cutting. In the disclosed embodiments, sensor data and other input data may be examined to determine physical characteristics of the build part, such as porosity and surface finish, and these sensor spaces may be used in the model to achieve a part of a desired quality.

In the disclosed embodiments, given various inputs, such as sensor inputs and process parameters, the model may predict a quality score, which in turn may be used to determine whether the build part will be acceptable. If the predicted part quality is not acceptable, various measures may be taken to improve the manufacturing process. In other words, given a model, given a response map with sensors, given build data, and a scan file (e.g., CLI build file), the quality score generator may be used to predict whether a build is acceptable. If the quality score indicates that the build will be unacceptable, the ILC attempts to learn what is unacceptable (e.g., via a machine learning algorithm) and corrects the scan file for the part being built to make future builds more acceptable.

In general, there may be many different disturbances acting on the manufacturing process. Without interference, one can design an ideal set of scan parameters, such as laser power, speed, etc., and one would expect that each time the set of parameters (i.e., "recipe") is executed, the result will be a part with the desired characteristics. However, this does not happen because there is interference acting on the entire system, deviating the process from its nominal value. Some of these can be predictable disturbances, e.g., if one tries to build the same geometry, then the thermal conductivity is the same disturbance for each instance (i.e., each build). Similarly, if the same machine is being used and there is a problem in the optical system, the problem is known and calibration can be done for this purpose. On the other hand, there will be some random and therefore construction specific interference. This interference cannot be compensated for in a predictable manner. Iterative Learning Control (ILC) is used to learn from historical builds and correct in subsequent builds, which may be considered a "feed forward" control process. This is only possible for predictable disturbances — the algorithm learns what can be predicted and compensated for. On the other hand, for random interference, a feedback control procedure may be used.

The ILC has a control algorithm that receives a tracking error in a first loop, the tracking error being determined based on a set of reference (i.e., expected) mass scores compared to mass scores predicted based on sensor data measured during the build. Based on the tracking error, the algorithm updates the build file (i.e., the scan file or set of parameters) for use in the next build iteration. Alternatively, as described above, a set of reference sensor data values may be used as a basis for comparison in the tracking error loop. In this case, rather than converting the measured sensor data to a mass fraction (using a reference surface) and comparing the mass fraction to a desired mass fraction target, the measured sensor data is compared to a reference sensor data value in the tracking loop.

In the disclosed embodiment, the ILC receives an estimated error in the second loop, which is a comparison of the predicted sensor data value and the measured sensor data. The sensor data is predicted using a thermal model that is initially a nominal thermal model but is subsequently updated by an algorithm based on the estimation error. The thermal model receives a build parameter, such as a scan file, and based on the input, predicts a set of sensor data values. In the case of a perfect thermal model, the predicted sensor data values will correspond exactly to the measured sensor data. Since the nominal thermal model is not perfect, the actual sensor response is different from the predicted sensor response — this difference is the estimation error. The estimation error may be fed to a tracking filter that compares the predicted sensor data values (i.e., predicted by the thermal model) with the measured sensor data and updates the thermal model in a manner suitable to minimize the estimation error.

The two loops of the ILC described above, which may be referred to as an iterative learning control loop and a tracking filter loop, respectively, may operate independently (e.g., one at a time) or may operate in combination. Note that the terminology used herein describes the tracking error as being fed to the ILC and the estimation error as being fed to the tracking filter.

In the disclosed embodiment, the ILC loop and the tracking filter loop are used in combination to iteratively minimize both types of errors, i.e., tracking error and estimation error. Minimization of tracking error means that the predicted mass fraction of the build part should be moved from an unacceptable range to an acceptable/desirable range. Minimization of the estimation error means that the thermal model approaches high accuracy. Therefore, the predicted sensor values will closely match the measured sensor data, which will allow the ILC loop (which relies on the thermal model) to converge faster. As mentioned above, the tracking filter loop may be used in the disclosed embodiments without a tracking filter loop, since the nominal thermal model may be sufficiently accurate in practice. As described above, in conventional methods, non-systematically applying power corrections to an entire part or region of a part to optimize, for example, surface finish, can result in effects on other material properties (e.g., porosity). On the other hand, the use of ILC allows build parameters (e.g., power) to converge in a manner that may result in improved material properties without significant trade-offs for either material property.

As mentioned above, if the part being built has a difficult geometry, such as a protruding region, there will be a relatively large melt pool due to the reduced thermal conductivity in the region in question, i.e. less heat is conducted away from this region through the bulk material of the part, resulting in a reduced part quality. In this case, a relatively large power reduction may be required to compensate. In other words, the larger the puddle, the greater the power "delta" required. In some cases, increments may also be applied to the laser speed and focus to help compensate for larger puddles.

In a disclosed embodiment, the corrected laser power level may be iteratively determined according to the following equation, where i is an integer representing the number of iterations (i-th iteration), and k is₁And k₂Is an experimentally determined gain coefficient for sensor error and thermal error, respectively:

Corrected_Power_i＝Power_i+k₁*tracking_error_i+k₂*estimation_error_i

the tracking error (tracking _ error) in the above equation is determined based on a set of reference (i.e., expected) mass scores that are compared to mass scores predicted based on sensor data measured during the build. Alternatively, as described above, a set of reference sensor data values may be used as a basis for comparison in the tracking error loop. As the iterations converge, the tracking error approaches zero.

The estimation error (estimation _ error) in the above equation is determined based on a comparison of the predicted sensor data value (i.e., predicted by the thermal model) and the measured sensor data. As the iteration converges, the estimation error approaches zero.

In the disclosed embodiments, a sensor error map may be determined based on a difference between a reference intensity and a photodiode sensor intensity (i.e., measured sensor data). If the photodiode sensor Intensity is denoted "Intensity_i". By dividing the median power setting (i.e., nominal laser power set in build file) over all scan paths (i.e., hatched) by the median measured intensity (i.e., hatched)Measured sensor data) to give a conversion of power (e.g., in watts). Then the (point-by-point) sensor error is:

sensor_error_i＝(Ref_i-Intensity_i)*intensity_to_power_scaling

this calculated sensor error will correspond to a tracking error in an alternative embodiment in which a set of reference sensor data values is used as the basis of comparison in a tracking error loop (i.e., an embodiment in which measured sensor data is compared to reference sensor data values in a tracking loop, rather than using a reference surface to convert the measured sensor data to a quality score and compare the quality score to a desired quality score target). The sensor error obtained in this way is related to the photodiode intensity. The corrected power is then given by:

sensor_corrected_Power_i＝Power_i+sensor_error_i

in embodiments where it is not possible or desirable to make successively small adjustments to the laser power, the sensor correction power determined in this manner may be applied as a uniform power setting to the sub-segments of the scan path. As described above, the ILC loop is adapted to adjust corrected _ Power_iSo that the waste heat in the bulk material of the part is more evenly distributed throughout the build process. In some cases, the input power may be continuously adjusted on a very fine time scale. However, many small adjustments to the laser may be costly. Therefore, instead of correcting the power point-by-point, it is sometimes more practical to specify a piecewise (i.e., sub-piecewise) constant model. For this purpose, a sub-segmentation procedure is employed, in which a sub-segment is determined which corrects for relative stability in power. Then, the median of the correction powers along the sub-segments is set as the correction power in the segment.

In some cases, the sensor error, and therefore the correction power, tends to be large near the beginning of the segment. In this case, a moving average of the correction power may be used for better stability of the algorithm. For example, for a slave (x)₀,y₀) Each shadow of the beginningLine, when y-y₀|>0.1y, as long as | x-x₀|>100 μm, a new sub-segment can be started at (x, y). Once the correction power signal stabilizes, the x threshold is increased to, for example, 300 μm. An example of this sub-segmentation process is as follows. For each sensor reading, an associated heat input q is determined for the region (e.g., voxel) in which it is located_in。

For a thermal model-based correction, the thermal model can be introduced in the following equation for the intensity measured by a sensor, such as an Avalanche Photodiode (APD):

Intensity＝C₁*Laser Power+C₂*VF+C₃。

this equation can be rewritten to solve for power as follows:

thermal_power_i＝(Intensity_i-C₂*VF_i-C₃)/C₁

in these equations, C₁、C₂And C₃Is an experimentally determined coefficient, and VF is the volume fraction determined for a particular voxel during the calculation of the nominal (i.e., initial) thermal model (the variable thermal Power is used instead of Laser Power).

May be based on heat input q_inAnd the difference between the measured sensor intensities. Then the thermal error is:

thermal_error_i＝Ref_i*intensity_to_power_scaling-thermal_power_i

wherein Ref_iIs the same reference intensity used to calculate the sensor error.

The correction power of the thermal model may be calculated as

thermal_corrected_power_i＝Power_i+thermal_error_i

Using equations similar to those presented herein for the correction power based on tracking error and estimation error, the sensor-based and thermal model-based power corrections can be combined to form a model for obtaining the total correction power for the next construction:

corrected_power＝Power+k₁*sensor_error+k₂*thermal_error

wherein k is₁And k₂Is the gain parameter.

The next iteration is built using the correction power in a new build file (e.g., CLI file). In the disclosed embodiment, for subsequent builds (i.e., after the first power correction build), additional steps are used. Instead of applying the sub-segmentation discussed above to the previous construction that has already been sub-segmented, the sub-segmentation process is applied to a set of original hatchings (i.e., scan paths). Thus, the new sensor data is first registered with the previous build so that the relevant input power can be obtained at each point. These segments are then identified with a set of original hatchings (obtained from the first uncorrected build). After this, the steps presented above can be applied to obtain a new power correction build.

In the disclosed embodiment, a mass fraction is defined and used in a probabilistic sense to indicate which defects have occurred (or are about to occur) in a built part. For example, a determined weld puddle temperature and size may be required for acceptable part quality for a particular process. Other techniques consider variables of interest, such as laser power and speed, and attempt to detect changes in these variables. However, such techniques do not provide an indication of whether the determined deviation would result in, for example, unfused, porosity, and cracks in the build part, as well as the likelihood of such defects. Mass fraction analysis, on the other hand, takes input data and predicts defects and their likelihood in the built part based on a mapping to mass fractions. Conventionally, such an evaluation output is obtained by performing destructive analysis (e.g., cutting) on the part. Once the parameter set is obtained, the build is performed using the parameter set, the sample (i.e., part) is removed from the machine, and the cut of the part is performed. The defect score can be obtained from automated analysis under a microscope. When this is done for a set of points, for a particular set of input parameters, a response surface is generated. Based on the scores determined from the analysis of the cut layers, there may be good portions, bad portions, and portions of the gray area in between for each layer. These dots may be visually depicted as having responsive surfaces with, for example, red, blue, and green regions. The response surface is a mathematical function derived experimentally with a set of input parameters, e.g. laser power, focus, scanning speed, hatch spacing and layer thickness. The output of the response surface (i.e., function) may be, for example, a porosity fraction, a crack (i.e., defect) fraction, and an unfused fraction.

The response surface may be represented as an N-dimensional vector function. For example, if a function has inputs X and Y and outputs A and B, then this function from { X, Y } to { A, B } will be referred to as a 2X2 vector function. In the disclosed embodiment, the response surface is an n × m vector function. The inputs may be build parameters such as power, speed, focus, hatch spacing, layer thickness, and measured reference sensor data. The output of the reference surface is a mass fraction derived from a physical evaluation of the material properties of the reference part, e.g. porosity, unfused, microcracks, etc. mapped on the surface and/or volume of the part.

A look-up table may be used to represent the reference surface, in which case a set of inputs looked up in the table will provide a set of predicted output values, e.g. quality scores. Computing the reference surface may require a large amount of data, but techniques such as extrapolation using machine learning techniques may enhance the measurement data. In the disclosed embodiments, a relatively simple reference surface that maps the porosity mass fraction to the laser power level may be used. Based on this simple reference surface, the collected sensor data can be input into a reference surface look-up table, which will output a quality score indicating whether the quality of the build part is acceptable. If the quality score determined in this manner is acceptable, the design, i.e., the build file, may be frozen and declared ready for production. If the quality score is not acceptable, an Iterative Learning Control (ILC) may be used to adjust parameters in the build file, as discussed in further detail below. Alternatively, a set of reference (i.e., target) sensor data values may be derived using the reference surface and used as a basis for comparison.

The disclosed embodiments provide for recording sufficient sensor data to extend the input space such that the sensor data, in conjunction with the set of input parameters, provides sufficient information to allow prediction of a defect score. Destructive testing was performed to generate x-mat surfaces, and these results were then generalized to apply to any build part. In other words, the experimentally obtained x-mat surface is aligned with the input parameters to obtain the system model.

Fig. 2 is a diagram of a system 200 for determining a quality score of a produced part using a quality score generator 140, the quality score generator 140 receiving sensor data from an additive manufacturing machine 110. The sensor data may include data from actuator sensors 210 (e.g., galvanometer position sensors) associated with actuators 220 in AMM 110. The sensor data may also include data from environmental sensors 230, such as, for example, atmospheric pressure, oxygen levels, airflow, smoke, and the like. The sensor data may also include data from sensors 240 (e.g., photodiodes, pyrometers, acoustic emissions, etc.) that monitor the characteristics of the molten bath. In addition, as explained in further detail below, the quality score generator 140 receives training data, such as a response surface/map 250 derived from experimental testing of the build part.

When the additive manufacturing machine 110 (e.g., a DMLM machine) is building, the machine generates output data from the melt pool sensor 240, for example in the form of a data file in technical data management flow (TDMS) format. Data is also generated by other sensors of the machine, such as actuator sensors 210 that measure the position of a galvanometer that positions the laser spot, and various environmental sensors. Other data sources include commands to the additive machine, material property data, and response surface/map 250. Build part characteristics are determined based on these various process parameters. As described above, data from the DMLM machine may be used as an input to a quality score generator, which outputs a quality score. In a simple case, such a quality score may be a "pass/fail" score. In the disclosed embodiment, the numerical score is used to represent the quality of the built part. In the disclosed embodiments, the quality score is used after (or during) the part is built to evaluate whether the part is of acceptable quality. This is done instead of or in conjunction with other more time consuming evaluation processes, such as cutting and analysis of the part.

Conventionally, such an evaluation output is obtained by performing a destructive analysis of the part, for example, a response map may be generated when the process is completed for a particular set of input parameters. Based on the scores determined from the analysis of the cut layers, there may be combinations of input parameters that produce good results, combinations that produce poor results, and gray regions in between. These combinations of input parameters may be visually depicted in the response map as axes of a 2D or 3D map, while the output (e.g., density or quality score of anomalies) may be represented by color-coded (e.g., red, blue, and green) regions.

In the disclosed embodiments, the response map (e.g., the "response surface") may be a direct illustration of experimental data on 2D, 3D, or 3D, for example, with a color-coded map, or a mathematical function derived through experimentation, with inputs given by: (i) a set of parameters such as laser power, focus offset or beam spot size, scanning speed, hatch spacing and layer thickness, and/or (ii) measured or derived process variables such as bath depth, bath width, bath temperature and/or thermal gradients. The output obtained from the response map may be a color-coded map of, for example, the density of anomalies or defects, such as the area or volume percentage of voids, cracks, and unfused defects. It should be noted that the term "response surface" is used to describe a mathematical relationship between various process inputs and the density of anomalies, such as described above, rather than things related to the physical surface of the part being built.

The disclosed embodiments provide for recording sufficient sensor data to expand the input space such that the sensor data, in conjunction with a set of input parameters (e.g., a build file), provides sufficient information to allow for prediction of defect scores (e.g., mass scores) without cutting (i.e., sectioning) and/or other direct part testing, such as Optical Coherence Tomography (OCT) imaging. In the disclosed embodiments, experiments (e.g., physical tests) are conducted to generate response surfaces, and then the results are generalized to apply to any build part, for example, by creating a model.

In the disclosed embodiments, associations are created between all input variables and some form of quantization concept of the quality scores, which may be discrete or continuous (e.g., low/medium/high or real numbers). An initial version of the model (e.g., a regression model) may be used to establish a direct correlation from the input variables to the output quality scores. Such a model may use an equation expressed in terms of input variables and coefficients, i.e., a regression model. The relationship between input variables and output can be highly non-linear and complex, as there can be a large number of inputs (e.g., intensity of each pixel of a 256x256 pixel image) and there can be only one output, i.e., a quality score. Transformations of the input variables may be created, i.e. explicitly transforming the input variables into a "feature space" or neural network, decision trees, etc. may be used, i.e. machine learning. This provides space to make the mapping problem easier. In other words, one can construct a latent variable space to simplify the problem, starting with the direct variable. In particular, machine learning can be used to acquire and map a high-dimensional multivariate space to an output, where the underlying relationships are known to be complex, non-linear, and non-trivial.

Fig. 3 is a diagram of the quality score generator 140. As described above, the quality score generator 140 receives sensor data 130 from the AMM, a nominal build file 120 of the part being produced, and reference data 250 derived from testing of the built part. Sensor data 130 and nominal build file 120 are input to machine learning algorithm 310, and machine learning algorithm 310 is trained to generate a quality score for the build part. The machine learning algorithm 310 is trained using the response surface/map 250, which, as described above, the response surface/map 250 is derived from physical testing of the part. Such testing may involve slicing the build part and analyzing each slice, for example using Optical Coherence Tomography (OCT), to determine the presence and density of various types of anomalies. A quality score calculation 330 may be performed, for example, combining the quality scores of different types of anomalies, and the results output by the quality score generator 140. Alternatively, the machine learning algorithm 310 may determine the quality score that is directly output by the quality score generator 140.

In the disclosed embodiment, three types of anomalies may be considered: porosity, cracks and unfused defects. An indication of the overall area or volume percentage and/or density of such anomalies, such as mass fractions, may be predicted for each of such features, or, depending on the relative importance of these features with respect to the desired physical and mechanical properties, these mass fractions may be combined to obtain a sum, maximum, weighted average, etc. For example, in some cases, cracks may be the most important features, while in other cases, pore density and unfused anomalies may be more important.

To train the machine learning algorithm 310, a cut of the build part may be made to produce the response surface/map 250. In the disclosed embodiment, the cut image may be divided into smaller sub-regions, such as regions of 3x3 pixel space (kxk, where k may be considered a parameter in general), to convert the image into a vector, even if the image is flattened. A numerical matrix may be generated with multiple inputs, for example, nine variables for each 3x3 pixel space, and one output variable. A determination is made whether the post-cut inspection reveals any anomalies in this 3x3 pixel space, meaning that one is looking at the image locally and asking if there is unfused or any porosity problems or other anomalies. The label is then assigned to the 3x3 pixel being examined. In other words, on a binary scale, this 3x3 pixel area has an anomaly. This is equivalent to a binary classification problem using the typical data format of the input as a machine learning model, although multiple classes versions of this problem (whereby different classes will be different defect types) can also solve multiple classes of classification problems using machine learning methods. In either case, the multivariate latent variable model, i.e., the machine learning model, can perform a mapping between a nine-element vector (typically, an n-element vector) to a single value. Using such a model, any 3x3 (typically, k x k) pixel combination of intensities can be created and fed into the model, which would indicate the likelihood of a defect (or defect type) being present in the corresponding sub-region of the build part.

In an alternative embodiment, instead of flattening the measured part data into a matrix, the image as a whole may be used by a machine learning algorithm 310, and the machine learning algorithm 310 may be, for example, a deep learning model such as a full convolution network or "U-Nets". Such a model can be used to construct predicted microscopic images directly from sensor data. In an alternative embodiment, a set of three-dimensional slices may be used instead of a two-dimensional image. In other words, one can examine the 3x3x3 pixel cube, rather than the 3x3 pixel group. Furthermore, although metal cross-sections have been described, it is also possible to generate a three-dimensional reconstructed volume from 2D Computed Tomography (CT) slices and to correlate the sensor data in 3D space with the 3D CT image.

In the disclosed embodiment, a statistical mass transfer function is developed to predict the density of particular anomalies in a build part. Various types of anomalies may be considered, such as voids, unfused defects, and cracks. Important parameters for the part being built may include the average value of the photodiode signal and certain process parameters, such as laser power setting and power divided by the laser scan speed. A linear or non-linear model may be used to provide the transfer function, which in the disclosed embodiment has a relatively high r-squared factor, e.g., a r-squared factor above about 0.8.

In the disclosed embodiment, there may be at least two types of response surfaces/graphs 250. The first type may be generated based on controlled experiments that attempt to characterize the material of the part based on input parameters (e.g., laser power, focus, and speed). In this case, the part may be produced and subjected to analysis, such as cutting and imaging. Algorithms, such as machine learning, may be used in conjunction with a relatively small number of iterations. The results of such experiments provide an indication of the area in the laser parameter space that gives the part a sufficiently low anomaly density. This in turn can be used to set the initial settings of an Additive Manufacturing Machine (AMM).

The second type of response surface/map 250 may include laser parameters such as those mentioned above in conjunction with sensor output data. For example, sensors such as photodiodes and cameras may be used to measure characteristics of the melt pool, e.g., size and temperature, while the manufacturing process is running. For example, the sensor data may indicate that the laser parameters do not necessarily translate into stable puddle characteristics. For example, the measured photodiode signal may not be constant, i.e., it may have variations and may not be a clean signal with respect to the spatial position of the part. Thus, a characteristic of the sensor output, e.g., the photodiode output signal, may provide another way to predict part quality. Thus, the information of the material properties that can be used to set the laser parameters provided by the first type of responsive surface can be supplemented by sensor readings to provide a more accurate model of part quality.

In the disclosed embodiment, the quality score generator 140 receives the sensor data 130 and applies a multidimensional mathematical formula or algorithm, such as a machine learning algorithm 310, to produce a quality score, which may be a number or a set of numbers. The algorithm 310 may be trained by performing several builds of the part and performing physical tests (e.g., cutting and/or volumetric CT, etc.) to measure anomalies/defects. This may include building a relatively simple reference part and building the part using different sets of laser parameters. Such experiments can be used to generate a response surface as an aid to the above-described experiments. The mass fraction generator may be adapted to use a formula that takes various types of anomalies, such as porosity, unfused, cracks, and combines the corresponding individual mass fractions to produce an overall mass fraction (e.g., by using a weighted average). The combined quality score may be adapted to give greater weight to a particular type of anomaly. Thus, the quality score algorithm may be trained experimentally, for example, through multiple iterations of producing and physically analyzing the part.

Once algorithm 310 is sufficiently trained, one can input measured sensor data 130 and nominal build file 110 for non-experimental cases, and algorithm 310 can output a response surface (e.g., a graph representing a multi-dimensional relationship between inputs such as laser parameters and sensor data and outputs such as anomaly densities) as if physical testing and analysis had been performed. The generated response surface may then be quantified according to the mass fraction. For example, the quality score may be obtained via further calculations (e.g., averaging of anomaly densities). Alternatively, algorithm 310 may directly output one or more quality scores, which may be used alone or mathematically combined. The determined quality score may be fed back 320 to the algorithm 310.

Fig. 4 depicts a system 400 for updating a thermal model 150 of a part 155 produced in an additive manufacturing process. As described above, the thermal model 150 uses the nominal build file 120 and sensor data 130 received from the AMM110 (e.g., a DMLM printer) to predict the thermal input (i.e., energy density) in the volume of the build part 155 that would result from applying a particular level of laser power during a scan. The thermal model 150 converts the predicted energy density into a corresponding predicted sensor reading, e.g., photodiode sensor data 410. Tracking filter 420 compares the predicted sensor readings 410 with actual measured sensor data 430 received from AMM 110. The differences are used to generate an updated thermal model 440, the updated thermal model 440 being output by the tracking filter 420 to replace the existing thermal model 150.

Fig. 5A depicts a portion of a part 500 that is voxelized to produce a nominal thermal model. For example, the part 500 may be divided into cubes, for example, having dimensions of about 0.3 to 1.0 mm. Other sizes are also possible and may be selected based on the geometry of the part 500. As the laser scans each voxel during the build process, a portion of the heat generated in the voxel dissipates to adjacent voxels 520. As shown in fig. 5B, for a particular voxel 510, the heat dissipation may be estimated by determining the number of neighboring voxels within a determined radius (r). For example, if voxel 510 is surrounded on all sides (as viewed from the top) by neighboring voxels 520, then voxel 510 may be expected to dissipate a relatively large portion of the heat generated by the scan. On the other hand, as shown in FIG. 5C, if voxel 510 (viewed from the side) is located at the top of the build part 500, the heat dissipation radius only applies to adjacent voxels 520 at or below the top surface, which leaves the entire (i.e., upper) hemisphere free of adjacent voxels to provide heat dissipation.

In the disclosed embodiment, a tracking loop is provided that starts with a nominal thermal model (i.e., a heat dissipation model for the part being built). In this case, the inputs to the ILC include the thermal model of the part and the scan file (e.g., CLI build file). Based on this, the ILC predicts what the sensor response will be based on thermal characteristics, e.g., which points in the part will be hotter than allowable, colder than allowable, etc. For example, if the same amount of energy is applied, the corners will therefore become hotter than the rest of the part with less heat flow/conductivity. On the other hand, in the middle region of the part there is much thermal conductivity (i.e. more paths for heat dissipation), so if the same amount of energy is applied, the problematic area will be cooler, since the heat may flow away more easily.

If a perfect thermal model is available, no iterative learning loop would be needed. In this case, one would have models and references, such as response surfaces/maps based on sensor data, so an ideal build file can be generated to achieve a certain defined quality result. In practice, an approximate model is available, which is sufficient for control purposes, but results in actual sensor data being different from predicted values. These differences (i.e., estimation errors) may be fed back through the tracking filter to update the model. Thus, for each build, the nominal thermal model is updated based on the estimation error. After several iterations, a nominal thermal model with an updated set of parameters will have very high fidelity, which will help the ILC to converge faster.

As described above, the determination of the nominal (i.e., initial) thermal model may include analyzing the part geometry, such as a part having a trapezoidal geometry in three dimensions, by creating voxels at different resolutions and calculating for each voxel how much metal is connected underneath it. For example, in the case of voxels in the middle of the part, then all voxel volumes will have solid metal below them, so the Volume Fraction (VF) will be 1. On the other hand, near the corners, the volume fraction may be about 1/4. In the case of the arch geometry, the voxels at the intermediate points are almost free of metal, so the volume fraction will be close to zero. In this way, the geometry is analyzed and a volume fraction is assigned to each voxel and the model is stored, for example in file format HDF 5. A set of experiments may be performed setting laser power, speed, and volume fraction of interest, and using the measured intensity sensor data to create a regression analysis that can predict the sensor data based on the input to the model and the volume fraction. This is a nominal thermal model of the predicted sensor data. If the thermal model does not match the measured intensity sensor data, the coefficients of the model (e.g., C0, C1, and C2, discussed herein) are adjusted based on a tracking filter algorithm to better match.

Fig. 6 is a block diagram of an apparatus 600 according to some embodiments. The apparatus 600 may comprise a general-purpose or special-purpose computing apparatus and may execute program code to perform any of the functions described herein. Apparatus 600 may include an implementation of one or more elements of system 100. According to some embodiments, the apparatus 600 may include additional elements not shown.

The apparatus 600 includes a processor 610 operatively coupled to a communication device 620, a data storage device/memory 630, one or more input devices (not shown), and one or more output devices 630. Network interface 620 may facilitate communication with external devices, such as application servers. The input device may be implemented in the apparatus 600 or in a client device connected via the network interface 620. Input devices may include, for example, a keyboard, buttons, a mouse or other pointing device, a microphone, knobs or switches, an Infrared (IR) port, a docking station, and/or a touch screen. The input devices may be used, for example, to manipulate a graphical user interface and enter information into apparatus 600. Output device 630 may include, for example, a display (e.g., a display screen), speakers, and/or a printer.

The data storage device/memory 640 may comprise any device, including a combination of magnetic storage devices (e.g., tape, hard drive, and flash memory), optical storage devices, Read Only Memory (ROM) devices, random access memory (memory), and the like.

The storage device 640 stores programs and/or platform logic for controlling the processor 610. The processor 610 executes instructions of a program to operate according to any of the embodiments described herein, including but not limited to processes.

The program may be stored in a compressed, uncompiled and/or encrypted format. Programs may also include other program elements, such as an operating system, a database management system, and/or device drivers used by processor 610 to interface with peripheral devices.

The foregoing figures represent logical architectures for describing processes according to some embodiments, and an actual implementation may include more or different components arranged in other ways. Other topologies may be used in conjunction with other embodiments. Moreover, each system described herein may be implemented by any number of computing devices in communication with each other via any number of other public and/or private networks. Two or more such computing devices may be remote from each other and may communicate with each other via any known network means and/or dedicated connection. Each computing device may include any number of hardware and/or software elements suitable for providing the functionality described herein, as well as any other functionality. For example, any computing device used in the implementation of system 100 may include a processor for executing program code, such that the computing device operates as described herein.

All of the systems and processes discussed herein can be embodied in program code stored on one or more computer readable non-transitory media. Such media, non-transitory media, may include, for example, fixed disks, floppy disks, CD-ROMs, DVD-ROMs, flash drives, magnetic tape, and solid state RAM or ROM storage units. Thus, embodiments are not limited to any specific combination of hardware and software.

Further aspects of the invention are provided by the subject matter of the following clauses:

1. a method for providing updated build parameters to an additive manufacturing machine, the method comprising: receiving sensor data from the additive manufacturing machine via a communication interface of a device comprising a processor during manufacturing of a part using a first set of build parameters; receiving the first set of build parameters; determining, using the processor of the device, an evaluation parameter based on the first set of build parameters and the received sensor data; generating, using the processor of the apparatus, thermal data based on a thermal model of the part derived from the first set of build parameters; applying, using the processor of the device, a first algorithm to the received sensor data, the determined evaluation parameters, and the generated thermal data to produce a second set of build parameters, the first algorithm being trained to improve the evaluation parameters; and outputting the second set of build parameters to the additive manufacturing machine to produce a second part.

2. The method of any of the preceding clauses wherein the evaluation parameter comprises a quality score determined by applying a second algorithm to the first set of build parameters and the received sensor data.

3. The method of any of the preceding clauses wherein the second algorithm is trained by receiving a reference derived from physical measurements made on at least one reference part constructed using a set of reference construction parameters.

4. The method of any of the preceding clauses, wherein the generating of the thermal data comprises calculating a first set of thermal data values based on a nominal thermal model and the first set of build parameters.

5. The method of any of the preceding clauses wherein the generating of the thermal data further comprises: determining an updated thermal model based on a comparison of the calculated first set of thermal data values and the received sensor data; and calculating a second set of thermal data values based on the updated thermal model.

6. The method of any of the preceding clauses wherein the nominal thermal model is derived by: dividing the volume of the part into voxels; determining the relative amount of surrounding material within a defined radius of the center of each of the voxels; and calculating a thermal data value for each voxel based on the relative amounts of surrounding material.

7. The method of any of the preceding clauses wherein the sensor data is received from at least one of a laser power sensor, an actuator sensor, a weld puddle sensor, and an environmental sensor.

8. A system for providing updated build parameters to an additive manufacturing machine, the system comprising: an apparatus comprising a communication interface, the apparatus configured to receive sensor data from the additive manufacturing machine during manufacturing of a part using a first set of build parameters, the apparatus further comprising a processor configured to: receiving the first set of build parameters; determining an evaluation parameter based on the first set of build parameters and the received sensor data; generating thermal data based on a thermal model of the part derived from the first set of build parameters; applying a first algorithm to the received sensor data, the determined evaluation parameters, and the generated thermal data to produce a second set of build parameters, the first algorithm being trained to improve the evaluation parameters; and outputting the second set of build parameters to the additive manufacturing machine to produce a second part.

9. The system of any of the preceding clauses wherein the evaluation parameter comprises a quality score determined by applying a second algorithm to the first set of build parameters and the received sensor data.

10. The system of any of the preceding clauses wherein the second algorithm is trained by receiving a reference derived from physical measurements made on at least one reference part constructed using a set of reference construction parameters.

11. The system of any of the preceding clauses wherein the generating of the thermal data comprises calculating a first set of thermal data values based on a nominal thermal model and the first set of build parameters.

12. The system of any of the preceding clauses wherein the generating of the thermal data further comprises: determining an updated thermal model based on a comparison of the calculated first set of thermal data values and the received sensor data; and calculating a second set of thermal data values based on the updated thermal model.

13. The system of any of the preceding clauses wherein the nominal thermal model is derived by: dividing the volume of the part into voxels; determining the relative amount of surrounding material within a defined radius of the center of each of the voxels; and calculating a thermal data value for each voxel based on the relative amounts of surrounding material.

14. A non-transitory computer readable storage medium storing program instructions that, when executed, cause a processor to perform a method for providing updated build parameters to an additive manufacturing machine, the method comprising: receiving sensor data from the additive manufacturing machine via a communication interface of a device including the processor during manufacturing of a part using a first set of build parameters; receiving the first set of build parameters; determining, using the processor of the device, an evaluation parameter based on the first set of build parameters and the received sensor data; generating, using the processor of the apparatus, thermal data based on a thermal model of the part derived from the first set of build parameters; applying, using the processor of the device, a first algorithm to the received sensor data, the determined evaluation parameters, and the generated thermal data to produce a second set of build parameters, the first algorithm being trained to improve the evaluation parameters; and outputting the second set of build parameters to the additive manufacturing machine to produce a second part.

15. The computer-readable storage medium of any of the preceding clauses, wherein the evaluation parameter comprises a quality score determined by applying a second algorithm to the first set of build parameters and the received sensor data.

16. The computer-readable storage medium of any of the preceding clauses wherein the second algorithm is trained by receiving a reference derived from physical measurements made on at least one reference part, the at least one reference part being constructed using a set of reference construction parameters.

17. The computer-readable storage medium of any of the preceding clauses, wherein the generating of the thermal data comprises calculating a first set of thermal data values based on a nominal thermal model and the first set of build parameters.

18. The computer-readable storage medium of any of the preceding clauses, wherein the generating of the thermal data further comprises: determining an updated thermal model based on a comparison of the calculated first set of thermal data values and the received sensor data; and calculating a second set of thermal data values based on the updated thermal model.

19. The computer-readable storage medium of any of the preceding clauses wherein the nominal thermal model is derived by: dividing the volume of the part into voxels; determining the relative amount of surrounding material within a defined radius of the center of each of the voxels; and calculating a thermal data value for each voxel based on the relative amounts of surrounding material.

20. The computer readable storage medium of any of the preceding clauses wherein the sensor data is received from at least one of a laser power sensor, an actuator sensor, a weld puddle sensor, and an environmental sensor.

The embodiments described herein are for illustrative purposes only. Those skilled in the art will recognize that other embodiments may be practiced with modifications and alterations to the above-described embodiments.