Control system for adaptive control of heat treatment system

Abstract

A method for controlling a heat treatment system is provided, including: setting a temperature setpoint curve by a control system of the heat treatment system; determining, by the control system, an actual workpiece temperature estimate associated with the workpiece; comparing, by the control system, the actual workpiece temperature to one of a plurality of temperatures in a temperature setpoint curve; determining, by the control system, a difference between the actual workpiece temperature estimate and the temperature setpoint curve; adjusting, by the control system, an operating parameter of the heat treatment system based at least in part on a difference between the actual workpiece temperature estimate and the temperature setpoint curve; tracking, by the control system, an actual workpiece temperature specified by the temperature setpoint curve with respect to the actual workpiece temperature estimate; and accessing, by the control system, a system model providing a simulated temperature estimate associated with the workpiece, the simulated temperature estimate simulating an actual workpiece temperature estimate associated with the workpiece, the system model including a machine learning model.

Description

[0001] This application is a divisional application of Chinese application No. 202080091407.3, filed on October 29, 2020, entitled “Control system for adaptive control of thermal processing system”. TECHNICAL FIELD

[0002] The present disclosure generally relates to thermal processing systems. BACKGROUND

[0003] As used herein, a thermal processing chamber refers to an apparatus that heats a workpiece, such as a semiconductor wafer. Such an apparatus can include a support plate for supporting one or more semiconductor wafers and an energy source, such as a heating lamp, laser, or other heat source, for heating the semiconductor wafers. During thermal processing, the semiconductor wafers can be heated under controlled conditions according to a pre-set temperature profile.

[0004] Many semiconductor heating processes require heating wafers to high temperatures so that various chemical and physical transformations can occur as the wafers are fabricated into device(s). For example, during rapid thermal processing, a semiconductor wafer can be heated by an array of lamps through a support plate to approximately 300 °C to approximately 1,200 °C for a duration typically less than a few minutes. The range and duration of these temperatures are typically determined by a temperature setpoint profile. In these processes, a primary goal can be to adaptively control the thermal processing system to improve temperature setpoint profile tracking performance. SUMMARY

[0005] Aspects and advantages of embodiments of the present disclosure will be set forth in part in the following description, or can be apparent from the description, or can be learned through practice of the embodiments.

[0006] One example aspect of the present disclosure relates to a method for controlling a thermal processing system, the method comprising: setting, by a control system of the thermal processing system, a temperature setpoint profile, wherein the temperature setpoint profile specifies a plurality of temperatures that can be applied to a workpiece; determining, by the control system of the thermal processing system, an actual workpiece temperature estimate associated with the workpiece; comparing, by the control system of the thermal processing system, the actual workpiece temperature to one of the plurality of temperatures in the temperature setpoint profile; determining, by the control system of the thermal processing system, a difference between the actual workpiece temperature estimate and the temperature setpoint profile; adjusting, by the control system of the thermal processing system, an operating parameter of the thermal processing system based at least in part on the difference between the actual workpiece temperature estimate and the temperature setpoint profile; tracking, by the control system of the thermal processing system, an actual workpiece temperature specified by the temperature setpoint profile with respect to the actual workpiece temperature estimate; and accessing, by the control system of the thermal processing system, a system model, wherein the system model provides a simulated temperature estimate associated with the workpiece, the simulated temperature estimate simulating the actual workpiece temperature estimate associated with the workpiece, the system model comprising a machine learning model.

[0007] Another example aspect of this disclosure relates to a heat treatment system comprising: a treatment chamber; a workpiece support operable to support a workpiece during heat treatment within the treatment chamber; one or more heat sources operable to heat the workpiece in the treatment chamber during heat treatment; one or more sensors configured to acquire data associated with a workpiece temperature; and a control system configured to perform operations for controlling the heat treatment system, the operations for controlling the heat treatment system including: determining an estimated actual workpiece temperature associated with the workpiece; comparing the actual workpiece temperature with one of a plurality of temperatures in a temperature setpoint curve; determining a difference between the estimated actual workpiece temperature and the temperature setpoint curve; adjusting operating parameters of the heat treatment system at least in part based on the difference between the estimated actual workpiece temperature and the temperature setpoint curve; tracking the actual workpiece temperature specified by the temperature setpoint curve relative to the estimated actual workpiece temperature; and accessing a system model, wherein the system model provides a simulated temperature estimate associated with the workpiece, the simulated temperature estimate simulating the estimated actual workpiece temperature associated with the workpiece, the system model including a machine learning model.

[0008] These and other features, aspects, and advantages of the various embodiments will be better understood with reference to the following description and the appended claims. The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments of the present disclosure and, together with the description, serve to explain the relevant principles. Attached Figure Description

[0009] A detailed discussion of embodiments for those skilled in the art is set forth in the description with reference to the accompanying drawings, wherein: Figure 1 An example Rapid Thermal Processing (RTP) system with a support plate having a spatially arranged low-transmission zone is described according to an example embodiment of the present disclosure. Figure 2 An example schematic representation of a control routine according to an example embodiment of the present disclosure is depicted, which is operable to use a learning routine to train a system model to determine workpiece temperature and other characteristics based on parameters of the heat treatment system. Figure 3 A schematic representation of a control routine according to an example embodiment of the present disclosure is shown, which is operable to train a workpiece deformation sub-model of a system model to simulate the radiation measurement output of a heat treatment system. Figure 4 A schematic representation of a control routine according to an example embodiment of the present disclosure is shown, which is operable to train a workpiece / cavity optical sub-model of a system model to simulate the optical output of a heat treatment system; Figure 5A schematic representation of a control routine according to an example embodiment of the present disclosure is shown, which is operable to train a workpiece / chamber thermodynamic model of a system model to simulate the heat output of a heat treatment system; Figure 6 A schematic representation of a control routine according to an example embodiment of the present disclosure is shown, which is operable to implement a control tuner to generate temperature setpoint tracking improvements for a heat treatment system during workpiece machining. Figure 7 An example representation depicts the result of applying one or more temperature setpoint tracking improvements to a heat treatment system according to an exemplary embodiment of the present disclosure; Figure 8 A flowchart depicts a process for adaptively controlling a heat treatment system according to an example embodiment of the present disclosure; and Figure 9 A flowchart is depicted for a process for adaptively controlling a heat treatment system according to an example embodiment of the present disclosure. Detailed Implementation

[0010] Reference will now be made in detail to embodiments illustrated in the accompanying drawings, which include one or more examples thereof. Each example is provided by way of explanation of the embodiments and not as a limitation thereof. Indeed, it will be apparent to those skilled in the art that various modifications and variations can be made to the embodiments without departing from the scope or spirit of this disclosure. For example, features shown or described as part of one embodiment may be used with another embodiment to produce yet another embodiment. Therefore, various aspects of this disclosure are intended to cover such modifications and variations.

[0011] An exemplary aspect of this disclosure relates to a control system for adaptively controlling heat applied to a workpiece in a heat treatment system. The workpiece may include, for example, a semiconductor workpiece, an optoelectronic workpiece, a flat panel display, or other suitable workpiece. The workpiece material may include, for example, silicon, silicon germanium, glass, plastic, or other suitable materials. In some embodiments, the workpiece may be a semiconductor wafer. Heat may be applied to these workpieces based on a temperature setpoint profile.

[0012] Temperature setpoint profiles can be configured to specify multiple different temperatures applied to a workpiece over time for heat treatment. For example, a temperature setpoint profile could specify that the workpiece is heated at a first temperature for 10 seconds, then raised to a second temperature and heated at the second temperature for less than 1 second. The temperature range and number of temperature setpoint profiles can be specified to occur at specific times. In some embodiments, temperature setpoint profiles can be specified or selected by a user or technician as part of a process formulation.

[0013] During processing, the control system for the heat treatment system is operable to determine an estimated actual workpiece temperature associated with the workpiece. This estimated actual workpiece temperature can be determined based on data from one or more sensors within the heat treatment system, including but not limited to radiometers, optical detectors, pyrometers, visible wavelength imaging systems, infrared imaging systems, or any other sensors operable to estimate the workpiece temperature.

[0014] The control system can compare the actual workpiece temperature estimate with the temperature specified by the temperature setpoint curve to determine the error or difference between the actual workpiece temperature estimate and the temperature setpoint curve. The control system can adjust the operating parameters of the heat treatment system (e.g., the heat emitted by the heat source, for example, by controlling the lamp power). According to an example aspect of this disclosure, the control system can implement control routines to improve the performance of the heat treatment system by accurately tracking the temperature specified by the temperature setpoint curve relative to the actual temperature estimate of the workpiece during heat treatment, thereby reducing the error.

[0015] In some embodiments, the control system can access a system model. The system model can be trained to simulate the conditions and outputs of the heat treatment system. More specifically, the system model can be trained to provide a simulated temperature estimate associated with the workpiece, based on specified operating conditions of the heat treatment system, simulating an actual temperature estimate or actual temperature output associated with the workpiece. Furthermore, the system model can be trained to provide simulated workpiece characteristics associated with the workpiece, simulating simulated characteristics of actual temperature, optics, radiation, emission, and other characteristics associated with the workpiece.

[0016] According to an example aspect of this disclosure, the system model can be a machine learning model. For example, the system model can be a random forest classifier; a logistic regression classifier; a support vector machine; one or more decision trees; a neural network; and / or other types of models including both linear and nonlinear models. The neural network can include deep neural networks, feedforward neural networks, recurrent neural networks (e.g., long short-term memory recurrent neural networks), convolutional neural networks, or combinations thereof. For example, the system model can be trained to provide simulated temperature estimates using a machine learning neural network trained by a learning routine using data collected from the machining of test and / or actual workpieces.

[0017] In some embodiments, a system learning routine can be used to train the system model. The error value can be determined at least in part based on the difference between the simulated temperature estimate and the actual workpiece temperature estimate. The system learning routine can be a differentiable objective function; when optimized, this function can be operated to provide modifications to the system model that reduce the error value associated with the difference between the simulated workpiece temperature and the actual temperature output obtained from the test workpiece, and / or the difference between the simulated workpiece characteristics associated with the workpiece and the actual temperature and optical characteristics associated with the workpiece. The system learning routine can receive one or more inputs, including but not limited to the difference between the actual system output and the simulated temperature estimate, and the difference between the simulated workpiece characteristics and the actual workpiece characteristics.

[0018] In some embodiments, a test workpiece can be used to obtain data associated with the workpiece temperature to train a system model for the heat treatment system. The data associated with the workpiece temperature obtained using the test workpiece can be accurate enough to be considered an actual temperature output. The test workpiece may include one or more sensors operable to measure the heat applied to the workpiece. For example, the test workpiece may include one or more thermocouples to measure the heat of the workpiece. Thermocouples can measure heat with sufficient accuracy to provide an actual temperature output. Furthermore, the optical characteristics associated with the test workpiece can be known. The temperature and optical characteristics associated with the test workpiece can be pre-measured and / or calculated to aid in testing and / or calibrating the performance of the heat treatment system.

[0019] In some embodiments, the system learning routine may receive a first temperature difference between an actual temperature output associated with the test workpiece and a simulated temperature estimate associated with the test workpiece. The system learning routine may also receive a first optical difference between an actual optical characteristic associated with the test workpiece and an estimated optical characteristic associated with the test workpiece. Based at least in part on the first temperature difference and / or the first optical difference, the system learning routine may generate one or more parameter modifications to modify one or more parameters associated with the system model.

[0020] For example, a system learning routine can generate a first simulated temperature estimate associated with the test workpiece, which differs significantly from a first actual temperature output associated with the test workpiece. The system learning routine can generate one or more parameter modifications to modify one or more parameters associated with the system model. The control system can apply one or more parameter modifications to the system model. The system model can then generate a second simulated temperature output associated with the test workpiece, which is more accurate than the first simulated temperature output.

[0021] In some embodiments, the system model may include multiple sub-models. System model sub-models can more effectively simulate a subsystem of the system. For example, the system model may include a workpiece deformation sub-model to more accurately simulate the effects of workpiece deformation in the system. As another example, the system model may include a workpiece / cavity optics sub-model to more accurately simulate the optical properties of the workpiece. And yet another example, the system model may include a workpiece / cavity thermal sub-model to more accurately simulate and estimate the workpiece temperature.

[0022] In some embodiments, the control system may use optimization methods to modify the parameters of the system model. The optimization method may be one or more of backpropagation, stochastic gradient descent, mini-batch gradient descent, or any other first- or second-order optimization method applicable to machine learning models. For example, stochastic gradient descent may be used to apply one or more modifications to the system model. As another example, backpropagation combined with mini-batch gradient descent may be used to determine and apply one or more modifications. Yet another example, backpropagation combined with different first-order optimization methods may be used to determine and apply one or more modifications.

[0023] In some embodiments, the control system may include a system controller. The system controller may be configured to control one or more operating parameters of the heat treatment system. Operating parameters of the heat treatment system may include, but are not limited to, heat output, duration and / or heat intensity, or any other adjustments to any component of the heat treatment system. For example, the system controller may control one or more operating parameters of the heat treatment system to adjust the heat output from one or more components of the heat treatment system, such as lamps.

[0024] In some embodiments, the control system may include a reliable control tuner. The reliable control tuner can adjust one or more system controller parameters of the system controller. Adjusting one or more system controller parameters can at least partially affect how the system controller uses its output to control one or more operating parameters of the heat treatment system. One or more system controller parameters may be, for example, gains used in proportional-integral (PI) controllers, proportional-derivative (PD) controllers, and / or proportional-derivative-integral (PDI) controllers.

[0025] For example, a reliable control tuner can adjust one or more system controller parameters of a system controller. Based at least in part on adjustments to one or more system controller parameters, the system controller can adjust one or more operating parameters of the heat treatment system to use the system controller output to adjust the heat output from one or more components of the heat treatment system.

[0026] In some embodiments, the control system may include a clone system controller operable to provide an output indicative of controller parameters to a system model. Based at least in part on the clone system controller output from the clone system controller, the system model can provide a simulated workpiece temperature estimate. Initially, the clone system controller output may be a clone controller output that mirrors or clones the actual controller output of the control system for the heat treatment system. The clone system controller may include one or more clone controller parameters. These one or more clone controller parameters may be, for example, gains used in proportional-integral (PI) controllers, proportional-derivative (PD) controllers, and / or proportional-derivative-integral (PDI) controllers.

[0027] In some embodiments, the one or more clone controller parameters may be adjusted by a clone control tuner. The clone control tuner is operable to modify one or more clone controller parameters of the clone system controller based on control tuner adjustments learned using tuning learning routines. More specifically, the clone control tuner is operable to provide adjustments to one or more clone controller parameters to the clone system controller based at least in part on the difference between the simulated workpiece temperature estimate and the temperature setpoint.

[0028] In some embodiments, the clone control tuner can be a model (e.g., a machine learning model) that correlates errors with adjustments in controller parameters. For example, the clone control tuner can be a random forest classifier; a logistic regression classifier; a support vector machine; one or more decision trees; a neural network; and / or other types of models that include both linear and nonlinear models. The neural network can include deep neural networks, feedforward neural networks, recurrent neural networks (e.g., long short-term memory recurrent neural networks), convolutional neural networks, or combinations thereof. For example, the clone control tuner can be trained to provide adjustments to one or more clone controller parameters using an artificial neural network.

[0029] In some embodiments, the clone control tuner can be trained using a tuning learning routine. The tuning learning routine can determine the tuning of the control tuner based on a simulated temperature estimate of the workpiece during heat treatment. In some embodiments, the tuning learning routine can implement a differentiable objective function; when optimized, this function is operable to provide modifications to the parameters of the clone control tuner that reduce the difference between the simulated workpiece temperature estimate and the temperature setpoint. The tuning learning routine can receive one or more inputs, including but not limited to the difference between the simulated workpiece temperature estimate and the temperature setpoint.

[0030] For example, a tuning learning routine can receive a first temperature difference between a simulated workpiece temperature estimate associated with the workpiece and a temperature setpoint associated with the workpiece. Based at least in part on the first temperature difference, the tuning learning routine can generate and apply one or more control tuner adjustments to adjust one or more parameters of the clone control tuner. In conjunction, the clone control tuner can apply clone controller parameter adjustments to the clone system controller based at least in part on the control tuner adjustments of the tuning learning algorithm. These clone controller parameter adjustments are operable to modify the output of the system model provided by the clone system controller to the performance of the simulated heat treatment system, thereby reducing the difference between the workpiece temperature estimate and the temperature setpoint.

[0031] In some embodiments, the tuning learning routine may use optimization methods to determine the tuning of the control tuner. The optimization method may be one or more of backpropagation, stochastic gradient descent, mini-batch gradient descent, or any other first- or second-order optimization method. For example, stochastic gradient descent may be used to apply one or more tunings to a clone control tuner. For another example, backpropagation may be combined with mini-batch gradient descent to determine and apply one or more tunings. For yet another example, backpropagation may be combined with different first-order optimization methods to determine and apply one or more tunings.

[0032] In some embodiments, clone controller parameter adjustments generated by the clone control tuner can be recorded in a recipe type recorder. Each of one or more clone controller parameter adjustments can be linked to one or more workpiece types or process recipes in the recipe type recorder. The recipe type recorder can be a database, data structure, record, etc. The recipe type recorder can track the performance of simulated workpiece temperature estimates generated by the clone controller parameter adjustments.

[0033] A trigger condition is met if the cloned system controller controls the system model to produce a more accurate simulated workpiece temperature estimate than a previously stored simulated workpiece temperature estimate stored in the recipe-type recorder. If the trigger condition is met, the cloned control tuner can provide the same cloned controller parameter adjustments to the trusted control tuner. The trusted control tuner can generate one or more system controller parameter adjustments for the system controller, at least in part, based on one or more cloned controller parameter adjustments.

[0034] For example, a clone control tuner can generate one or more clone controller parameter adjustments determined to reduce the difference between the simulated workpiece temperature estimate and the temperature setpoint. The clone control tuner can then provide these one or more clone controller parameter adjustments to a trusted control tuner. The trusted control tuner can then generate one or more system controller parameter adjustments for the system controller, at least in part, based on the clone controller parameter adjustments. The system controller parameters can be adjusted to more accurately track the actual workpiece temperature via the temperature setpoint profile.

[0035] In some embodiments, the system evaluates whether adjustments to one or more system controller parameters improve performance. If it is determined that the actual workpiece temperature estimate improves performance, the trusted control tuner can retain one or more system controller parameter adjustments. If it is determined that the actual workpiece temperature estimate degrades performance, the control system can determine that an adjustment error has occurred. If an adjustment error has occurred, the trusted control tuner can discard one or more system controller parameter adjustments and alert the system model watchdog. The system model watchdog can determine that the system model should be retrained based on the reduced performance. Alternatively, the system model watchdog can apply system model modifications to the system model, and these modifications can be operated to correct the adjustment error.

[0036] For example, a trusted control tuner can apply adjustments to one or more system controller parameters to the system controller. Actual workpiece temperature estimates may be affected by performance degradation in tracking the temperature setpoint. In response, the trusted control tuner can discard one or more system controller parameter adjustments. Furthermore, a system model watchdog can determine when the system model should be retrained and initiate the retraining process.

[0037] The various aspects of this disclosure can achieve numerous technical effects and benefits. For example, the various aspects of this disclosure can continuously reduce errors in temperature setpoint tracking of the system, thereby increasing heat treatment efficiency and productivity. Furthermore, the various aspects of this disclosure can be used to generate more robust system models for heat treatment systems, thereby resulting in more accurate workpiece temperature estimations, etc.

[0038] Variations and modifications may be made to these exemplary embodiments of the present disclosure. As used in the specification, the singular forms “a,” “and,” and “the” include plural references unless the context clearly specifies otherwise. The use of “first,” “second,” “third,” and “fourth” is as an identifier and refers to the order of processing. For purposes of illustration and discussion, reference may be made to “substrate,” “wafer,” or “workpiece.” Those skilled in the art using the disclosure provided herein will understand that the exemplary aspects of this disclosure can be used with any suitable workpiece. The term “about” used with numerical values ​​means within 20% of the specified value.

[0039] Example embodiments of this disclosure will now be discussed in detail with reference to the accompanying drawings. Figure 1 An example Rapid Thermal Processing (RTP) system 100 with a support plate 120 having a spatially arranged low-transmission zone is depicted according to an exemplary embodiment of the present disclosure. As shown, the RTP system 100 includes an RTP chamber 105, a workpiece 110, a support plate 120, heat sources 130 and 140 (e.g., lamps), an air bearing 145, a pyrometer 165, a controller 175, a door 180, and a gas flow controller 185.

[0040] The workpiece 110 to be processed is supported in an RTP chamber 105 (e.g., a quartz RTP chamber) by a support plate 120. The support plate 120 is a workpiece support operable to support the workpiece 110 during heat treatment. The support plate 120 includes a rotatable base 135 and at least one support structure 115 extending from the rotatable base 135. The support structure describes a configuration that contacts and supports the workpiece during heat treatment. Examples of support structures may include one or more support pins, ring supports, or any other suitable support that contacts and supports the workpiece. Figure 1 As shown, the support structure 115 includes one or more support pins (only one is shown). The support structure 115 and the rotatable base 135 can transfer heat from the heat source 140 and absorb heat from the workpiece 110. In some embodiments, the support structure 115 and the rotatable base 135 may be made of quartz. As further described below, the rotatable base 135 rotates the workpiece 110 in a defined direction of rotation and a defined rotational speed.

[0041] A protective ring (not shown) is used to reduce edge effects from radiation emanating from one or more edges of workpiece 110. End plate 190 seals the chamber 105, and door 180 allows workpiece 110 to enter and, when closed, allows chamber 105 to seal and introduce process gas 125 into chamber 105. Two sets of heat sources (e.g., lamps or other suitable heat sources) 130 and 140 are shown on either side of workpiece 110, operable to heat the workpiece within the chamber.

[0042] The gas flow 150 can be an inert gas that does not react with the workpiece 110, or the gas flow 150 can be an active gas such as oxygen or nitrogen that reacts with the material of the workpiece 110 (e.g., a semiconductor wafer) to form a layer on the workpiece 110. In some embodiments, an electric current can flow through the atmosphere in the RTP system 100 to generate ions that react with or at the surface, and additional energy is imparted to the surface by bombarding the surface with high-energy ions.

[0043] Controller 175 controls the rotatable base 135 to rotate the workpiece 110. For example, controller 175 generates instructions defining the rotation direction and speed of the rotatable base 135 and controls the rotatable base 135 to rotate the workpiece 110 in the defined rotation direction and speed. The rotatable base 135 is supported by an air bearing 145. An airflow 150 impacting the rotatable base 135 causes the rotatable base 135 to rotate about axis 155.

[0044] Controller 175 is used to control heat sources 130 and 140. Controller 175 can also be used to control gas flow controller 185, gate 180, and / or temperature measurement system, here indicated as pyrometer 165. Controller 175 can be dynamically modified to improve temperature setpoint tracking performance, which will be discussed in detail in the following figure.

[0045] As used herein, a controller, control system, or one or more components of a control system may include one or more processors and one or more storage devices. One or more processors may be configured to execute computer-readable instructions stored in one or more storage devices to perform operations, such as any operations for controlling the heat treatment system described herein.

[0046] For the purposes of explanation and discussion, Figure 1 An example heat treatment system 100 is depicted. Those skilled in the art will understand using the disclosure provided herein that various aspects of this disclosure can be used with other heat treatment systems for workpieces without departing from the scope of this disclosure.

[0047] Figure 2 An example schematic representation of control routine 200 is depicted, which is operable to use a learning routine to train a system model, thereby determining workpiece temperature and other characteristics based on parameters of the heat treatment system.

[0048] System controller 200 (e.g., Figure 1 The controller 175 is operable to provide controller outputs to control components of the heat treatment system 204 during the processing of the test workpiece. The test workpiece may include one or more sensors operable to measure the temperature of the workpiece. For example, the test workpiece may include one or more thermocouples to measure the temperature of the workpiece. The one or more sensors on the test workpiece may record the temperature accurately enough to produce an actual system output 218. In addition, the test workpiece may include actual workpiece characteristics 208. The actual workpiece characteristics 208 may be pre-measured and / or calculated to aid in testing and / or calibrating the performance of the heat treatment system.

[0049] The system controller 200 is also operable to provide controller outputs to the system model 206. The system model 206 can be trained to simulate the conditions and outputs of a heat treatment system with the test workpiece 204. More specifically, the system model 206 can be trained to provide a simulated temperature estimate 216 for the test workpiece 204 simulating the actual system output 218. Furthermore, the system model 206 can be trained to provide simulated workpiece characteristics 212 associated with the test workpiece to simulate the actual workpiece characteristics 208.

[0050] In some embodiments, system model 206 may be a machine learning model. For example, system model 206 may be a random forest classifier; a logistic regression classifier; a support vector machine; one or more decision trees; a neural network; and / or other types of models including both linear and nonlinear models. Neural networks may include deep neural networks, feedforward neural networks, recurrent neural networks (e.g., long short-term memory recurrent neural networks), convolutional neural networks, or combinations thereof. For example, system model 206 may utilize a machine learning neural network trained by learning routines to provide a simulated temperature estimate 216.

[0051] The system learning routine 210 can be used to train the system model 206. The system learning routine 210 can be a differentiable objective function; when optimized, it can be operated to provide parameter modifications 214 to the system model 206, which reduce the difference between the simulated temperature estimate 216 and the actual system output 218 and / or the difference between the simulated workpiece characteristic 212 and the actual workpiece characteristic 208. The system learning routine 210 can receive one or more inputs, including but not limited to the difference between the actual system output 218 and the simulated temperature estimate 216, and the difference between the simulated workpiece characteristic 212 and the actual workpiece characteristic 208.

[0052] The control system 200 can use optimization methods to apply parameter modifications 214 to the system model 206. The optimization method can be, but is not limited to, backpropagation, stochastic gradient descent, mini-batch gradient descent, or any other first- or second-order optimization method applicable to machine learning models. For example, stochastic gradient descent can be used to apply one or more modifications to the system model. As another example, backpropagation combined with mini-batch gradient descent can be used to determine and apply one or more modifications. Yet another example, backpropagation combined with different first-order optimization methods can be used to determine and apply one or more modifications.

[0053] System model 206 may include one or more sub-models. Figure 3 An example schematic representation of a control routine 300 according to an exemplary embodiment of the present disclosure is depicted, which is operable to train a workpiece deformation sub-model of a system model to simulate the radiation measurement output of a heat treatment system.

[0054] The lamp power level 302 can be determined at least in part based on the temperature setpoint 602, such as Figure 3 As shown. The lamp power level 302, along with the known test workpiece temperature and optical properties 318, is sent as input to the model parameter tuner 312. The lamp power level 302 is also sent as input to the radiometer / chamber model 304.

[0055] The radiometer / chamber physical model 304 can simulate the conditions and output of the radiometer in the heat treatment system, at least in part, based on the lamp power level 302. The radiometer / chamber model 304 includes multiple model parameters. These model parameters can be adjusted by the model parameter tuner 312 using model parameter adjustment 316.

[0056] Similarly, the workpiece deformation model 306 can simulate the conditions and outputs of workpiece deformation during heat treatment. The workpiece deformation model also includes multiple model parameters operable and adjustable by the model parameter tuner 312. The workpiece deformation model 306, in conjunction with the radiometer / chamber model 304, can generate an estimated radiometer signal 320. The estimated radiometer signal 320 is operable to estimate the actual radiometer signal associated with the radiometer of the heat treatment system.

[0057] A model parameter tuner 312 can be trained to dynamically adjust the parameters of the radiometer / chamber model 304 and the workpiece deformation model 306 using model parameter adjustment 316. The model parameter tuner 312 can be a machine learning model. For example, the model parameter tuner 312 can be a random forest classifier; a logistic regression classifier; a support vector machine; one or more decision trees; a neural network; and / or other types of models including both linear and nonlinear models. The neural network can include deep neural networks, feedforward neural networks, recurrent neural networks (e.g., long short-term memory recurrent neural networks), convolutional neural networks, or combinations thereof. For example, the model parameter tuner 312 can be an artificial neural network trained to dynamically adjust the parameters of the radiometer / chamber model 304 and the workpiece deformation model 306 using model parameter adjustment 316.

[0058] The model parameter tuner 312 can be trained using a tuning learning routine 310. The tuning learning routine 310 can be a differentiable objective function; when optimized, it can operate to provide tuning parameter adjustments 314 for the model parameter tuner, which reduce the difference between the estimated radiometer signal 320 and the actual radiometer signal 308. The tuning learning routine 310 can receive one or more inputs, including but not limited to the difference between the estimated radiometer signal 320 and the actual radiometer signal 308.

[0059] Control routine 300 can use optimization methods to apply model parameter adjustments 316 to the radiometer / chamber model 304 and the workpiece deformation model 306. The optimization method can be, but is not limited to, backpropagation, stochastic gradient descent, mini-batch gradient descent, or any other first- or second-order optimization method applicable to machine learning models. For example, stochastic gradient descent can be used to apply one or more adjustments to the radiometer / chamber model 304 and the workpiece deformation model 306. As another example, backpropagation combined with mini-batch gradient descent can be used to determine and apply one or more adjustments. Yet another example, backpropagation combined with different first-order optimization methods can be used to determine and apply one or more adjustments.

[0060] Figure 4 An example schematic representation of a control routine 400 according to an exemplary embodiment of the present disclosure is depicted, which is operable to train a workpiece / cavity optical sub-model of a system model to simulate the optical output of a heat treatment system.

[0061] The actual radiometer signal 402 can be determined at least in part based on the temperature setpoint 602, such as Figure 4 As shown. The actual radiometer signal 402, along with the known test workpiece temperature and optical properties 418, is sent as input to the model parameter tuner 412. The actual radiometer signal 402 is also sent as input to the workpiece / cavity optical model 404.

[0062] The workpiece / cavity optical model 404 can simulate the conditions and outputs of the optical characteristics of the heat treatment system, at least in part, based on the actual radiometer signal 402. The workpiece / cavity optical model 404 includes multiple model parameters. These model parameters can be adjusted using model parameter tuning 416 via model parameter tuner 412. The workpiece / cavity optical model 404 can generate predicted workpiece optical characteristics 420. These predicted workpiece optical characteristics 420 are operable to predict known workpiece optical characteristics 420 associated with the test workpiece of the heat treatment system.

[0063] A model parameter tuner 412 can be trained to dynamically adjust the parameters of the workpiece / cavity optical model 404 using model parameter adjustment 416. The model parameter tuner 412 can be a machine learning model. For example, the model parameter tuner 412 can be a random forest classifier; a logistic regression classifier; a support vector machine; one or more decision trees; a neural network; and / or other types of models including both linear and nonlinear models. The neural network can include deep neural networks, feedforward neural networks, recurrent neural networks (e.g., long short-term memory recurrent neural networks), convolutional neural networks, or combinations thereof. For example, the model parameter tuner 412 can be an artificial neural network trained to dynamically adjust the parameters of the workpiece / cavity optical model 404 using model parameter adjustment 416.

[0064] The model parameter tuner 412 can be trained using a tuning learning routine 410. The tuning learning routine 410 can be a differentiable objective function; when optimized, it can operate to provide tuning parameter adjustments 414 for the model parameter tuner 412, which reduce the difference between the predicted workpiece optical property 420 and the known workpiece optical property 408. The tuning learning routine 410 can receive one or more inputs, including but not limited to the difference between the predicted workpiece optical property 420 and the known workpiece optical property 408.

[0065] Control routine 400 can use optimization methods to apply model parameter adjustments 416 to the workpiece / cavity optical model 404. The optimization method can be, but is not limited to, backpropagation, stochastic gradient descent, mini-batch gradient descent, or any other first- or second-order optimization method applicable to machine learning models. For example, stochastic gradient descent can be used to apply one or more adjustments to the workpiece / cavity optical model 404. As another example, backpropagation combined with mini-batch gradient descent can be used to determine and apply one or more adjustments. Yet another example, backpropagation combined with different first-order optimization methods can be used to determine and apply one or more adjustments.

[0066] Figure 5 An example schematic representation of a control routine 500 according to an example embodiment of the present disclosure is depicted, which is operable to train a workpiece / chamber thermodynamic model of a system model to simulate the heat output of a heat treatment system.

[0067] The lamp power level 502 can be determined at least in part based on the temperature setpoint 602, such as Figure 5 As shown. The lamp power level 502, along with the known test workpiece temperature and optical properties 518, is sent as input to the model parameter tuner 512. The lamp power level 502 is also sent as input to the workpiece / chamber thermal model 504.

[0068] The workpiece / chamber thermal model 504 can simulate the conditions and outputs of the thermal characteristics of the heat treatment system, at least in part, based on the lamp power level 502. The workpiece / chamber thermal model 504 includes multiple model parameters. These model parameters can be adjusted by the model parameter tuner 512 using model parameter adjustment 516. The workpiece / chamber thermal model 504 can generate an estimated workpiece temperature 520. The estimated workpiece temperature 520 is operable to predict a known workpiece temperature 520 associated with a test workpiece in the heat treatment system.

[0069] A model parameter tuner 512 can be trained to dynamically adjust the parameters of the workpiece / chamber thermal model 504 using model parameter tuning 516. The model parameter tuner 512 can be a machine learning model. For example, the model parameter tuner 512 can be a random forest classifier; a logistic regression classifier; a support vector machine; one or more decision trees; a neural network; and / or other types of models including both linear and nonlinear models. The neural network can include deep neural networks, feedforward neural networks, recurrent neural networks (e.g., long short-term memory recurrent neural networks), convolutional neural networks, or combinations thereof. For example, the model parameter tuner 512 can be an artificial neural network trained to dynamically adjust the parameters of the workpiece / chamber thermal model 504 using model parameter tuning 516.

[0070] The tuning learning routine 510 can be used to train the model parameter tuner 512. The tuning learning routine 510 can be a differentiable objective function; when optimized, it can operate to provide tuning parameter adjustments 514 for the model parameter tuner 512, which reduce the difference between the estimated workpiece temperature 520 and the known workpiece temperature 508. The tuning learning routine 510 can receive one or more inputs, including but not limited to the difference between the estimated workpiece temperature 520 and the known workpiece temperature 508.

[0071] Control routine 500 can use optimization methods to apply model parameter adjustments 516 to the workpiece / chamber thermal model 504. The optimization method can be, but is not limited to, backpropagation, stochastic gradient descent, mini-batch gradient descent, or any other first- or second-order optimization method applicable to machine learning models. For example, stochastic gradient descent can be used to apply one or more adjustments to the workpiece / chamber thermal model 504. As another example, backpropagation combined with mini-batch gradient descent can be used to determine and apply one or more adjustments. Yet another example, backpropagation combined with different first-order optimization methods can be used to determine and apply one or more adjustments.

[0072] Figure 6 An example schematic representation of a control routine 600 according to an exemplary aspect of this disclosure is depicted, which is operable to control a tuner to generate temperature setpoint tracking improvements for a heat treatment system during workpiece processing. The control routine 600 can determine improvements to controller performance based on the actual processing of the workpiece to improve temperature tracking via the temperature setpoint profile.

[0073] Temperature setpoint 602 may be a precise temperature value associated with a specific time during the heat treatment of a workpiece, defined, for example, by a temperature setpoint curve. Temperature setpoint 602 may be provided by recipe type 634. In some embodiments, recipe type 634 may be determined and / or selected by a user or technician. Temperature setpoint 602 may also be recorded in setpoint recorder 628. System controller 604 may use system controller output 640 to control one or more operating parameters of heat treatment system 606 (e.g., heat source output, duration and / or heat intensity, other adjustments to heat treatment system components, etc.). System controller output 640 may be based at least in part on temperature setpoint 602. Heat treatment system 606 may process workpieces. Data from one or more sensors in heat treatment system 606 may be processed to determine an actual workpiece temperature estimate associated with the workpiece in heat treatment system 606. Actual workpiece temperature estimate 608 may be based at least in part on system controller output 640.

[0074] For example, temperature setpoint 602 can specify a temperature of 500 degrees Celsius. System controller 604 can use system controller output 640 to control the operating parameters of heat treatment system 606 to heat the workpiece in heat treatment system 606 to 500 degrees Celsius. Data from sensors (e.g., pyrometers) in heat treatment system 606 can be processed to determine an actual workpiece temperature estimate 608 associated with the estimated temperature of the workpiece in heat treatment system 606.

[0075] The clone system controller 626 can use the clone system controller output 618 to output one or more operating parameters to the system model 616. The clone system controller output 618 may be based at least in part on a temperature setpoint 602. The clone system controller 626 may be a simulation of the system controller 604. For example, the clone system controller 626 may be operable to perform the clone system controller output 618, which is initially substantially similar to the system controller output 640.

[0076] According to an example embodiment of this disclosure, system model 616 may be a trained machine learning model to simulate heat treatment system 606. System model 616 may receive clone system controller output from clone system controller 626 in the same manner as heat treatment system 606 receives system controller output 640 from system controller 604. System model 616 is operable to model the conditions and outputs of heat treatment system 606. More specifically, system model 616 may be trained to provide a simulated workpiece temperature estimate 610 associated with a first workpiece of heat treatment system 606, which simulates an actual workpiece temperature estimate 608 associated with the first workpiece of heat treatment system 606.

[0077] In addition, the system model 616 can be trained to provide predicted workpiece characteristics associated with the workpiece, including but not limited to simulated temperature and optical characteristics that simulate the actual temperature and optical characteristics associated with the workpiece.

[0078] In some embodiments, system model 616 may be a machine learning model. For example, the system model may be a random forest classifier; a logistic regression classifier; a support vector machine; one or more decision trees; a neural network; and / or other types of models including both linear and nonlinear models. The neural network may include deep neural networks, feedforward neural networks, recurrent neural networks (e.g., long short-term memory recurrent neural networks), convolutional neural networks, or combinations thereof. For example, system model 616 may be trained to use an artificial neural network to provide a simulated workpiece temperature estimate.

[0079] The simulated workpiece temperature estimate 610, along with the actual workpiece temperature estimate 608, can initially be provided to the system model watchdog 612. The system model watchdog 612 can determine, at least in part, whether the system model 616 is sufficiently accurate based on the difference between the actual workpiece temperature estimate 608 and the simulated workpiece temperature estimate 610. If the system model watchdog 612 determines that the system model 616 is not accurate enough, the system model 616 may undergo a process similar to that previously... Figure 2 The retraining phase is basically similar to the process described in the text.

[0080] If the system model watchdog 612 determines that the system model 616 is accurate enough, then the system model 616 can provide the simulated workpiece temperature estimate 610 to the tuning learning algorithm 620 and the clone control tuner 622.

[0081] The clone control tuner 622 is operable to generate one or more clone controller parameter adjustments 630 for the clone system controller 626. More specifically, the clone control tuner 622 can be trained to provide one or more clone controller parameter adjustments 630 based at least in part on the difference between the simulated workpiece temperature estimate 610 and the temperature setpoint 602. The clone control tuner 622 can be a machine learning model. For example, the clone control tuner 622 can be a random forest classifier; a logistic regression classifier; a support vector machine; one or more decision trees; a neural network; and / or other types of models including both linear and nonlinear models. The neural network can include deep neural networks, feedforward neural networks, recurrent neural networks (e.g., long short-term memory recurrent neural networks), convolutional neural networks, or combinations thereof. For example, the clone control tuner 622 can be trained to use an artificial neural network to provide one or more clone controller parameter adjustments.

[0082] The clone control tuner 622 can be trained using a tuning learning routine 620. The tuning learning routine 620 can be a differentiable objective function; when optimized, it can operate to provide a control tuner adjustment 636 for the clone control tuner 622 that reduces the difference between the simulated workpiece temperature estimate 610 and the temperature setpoint 602. Alternatively or in combination, the clone control tuner 622 can provide a clone controller parameter adjustment 630 based at least in part on the difference between a predicted workpiece characteristic 632 and a known workpiece characteristic associated with a recipe type 634 stored in a recipe type recorder 624. The tuning learning routine 620 can receive one or more inputs, including but not limited to the difference between the simulated workpiece temperature estimate 610 and the temperature setpoint 602.

[0083] The control system 600 can use optimization methods to adjust the parameters of the cloned control tuner 622. The optimization method can be, but is not limited to, backpropagation, stochastic gradient descent, mini-batch gradient descent, or any other first- or second-order optimization method applicable to machine learning models. For example, stochastic gradient descent can be used to apply one or more control tuner adjustments 636 to the cloned control tuner 622. As another example, backpropagation combined with mini-batch gradient descent can be used to determine and apply one or more adjustments. Yet another example, backpropagation combined with different first-order optimization methods can be used to determine and apply one or more adjustments.

[0084] The clone controller parameter adjustments 636 generated by the clone control tuner 622 can be recorded in the recipe type recorder 634. The recipe type recorder 644 can track the performance of the simulated workpiece temperature estimate 610 generated by the clone controller parameter adjustments 630. If the clone system controller 626 uses the clone system controller output 618 to control the system model 616 to produce a more accurate simulated workpiece temperature estimate 610 than the simulated workpiece temperature estimate 610 previously stored in the recipe type recorder, the clone control tuner 622 can provide the same clone controller parameter adjustments 622 to the trusted control tuner 614.

[0085] The trusted control tuner 614 is operable to apply system controller parameter adjustment 638 to system controller 604. If the trusted control tuner 614 receives clone controller parameter adjustment 630 from clone control tuner 622, the trusted control tuner 614 can generate a system controller parameter adjustment 638 that is substantially similar to the clone controller parameter adjustment 630. The trusted control tuner 614 can apply system controller parameter adjustment 638 to system controller 604, and system controller 604 can control heat treatment system 606 at least in part based on system controller parameter adjustment 638.

[0086] For example, the clone control tuner 622 can generate one or more clone controller parameter adjustments 630, which are determined to reduce the difference between the simulated workpiece temperature estimate 610 and the temperature setpoint 602. The clone control tuner 622 can then provide these one or more controller parameter adjustments 630 to the trusted control tuner 614. The trusted control tuner 614 can generate one or more system controller parameter adjustments 638 that are substantially similar to the one or more clone controller parameter adjustments 630. The trusted control tuner 614 can apply the system controller parameter adjustments to the system controller 604, adjusting the way the system controller 604 controls the heat treatment system 606 using the system controller output 640.

[0087] Due to one or more system controller parameter modifications 638, the control system 600 can evaluate whether the actual workpiece temperature estimate 608 is more accurate. If it is determined that the actual workpiece temperature estimate 608 is more accurate, the trusted control tuner 614 will retain one or more system controller parameter modifications 638. If it is determined that the actual workpiece temperature estimate 608 is less accurate, the trusted control tuner 614 will discard one or more system controller parameter modifications 638 and warn the system model watchdog 612. The system model watchdog 612 can determine that the system model 616 should be retrained based on the reduced performance. This can be achieved through a process similar to... Figure 2 The system model 616 was retrained using the method described in the text.

[0088] Figure 7 An example representation 700 depicts the result of applying one or more temperature setpoint tracking improvements to a heat treatment system according to an exemplary embodiment of the present disclosure. As shown, representation 700 includes a temperature setpoint 702, an estimated actual workpiece temperature 704, an error example 706, and a correction example 708.

[0089] Indicated by 700, error example 706 is included, which demonstrates that in system controller 604, error example 706 has been... Figure 6 The reliable control tuner 614 adjusts the calculated difference between the previously existing temperature setpoint 702 and the actual workpiece temperature estimate 704. The temperature setpoint 702 can be obtained from the recipe and can be invoked to specify a particular temperature to be applied to the heat treatment system for a specific time. The actual workpiece temperature estimate 704 can be measured intermittently over time to track the temperature setpoint 702 as closely as possible. The accuracy of the actual workpiece temperature estimate 702 is based at least in part on the system controller output 640 of the system controller 604, such as... Figure 6 As shown.

[0090] Correction example 708 demonstrates a calculated reduction in the difference between the temperature setpoint 702 and the actual workpiece temperature estimate 704, as previously described in error example 706. This reduction can be at least partially attributed to adjustments 638 of one or more system controller parameters applied to system controller 604 by the reliable control tuner 614, such as... Figure 6 As shown in calibration example 708, the calibration process is iterative and can be operated continuously during heat treatment to continuously improve the tracking of the actual workpiece temperature estimate 704 to the temperature setpoint 602.

[0091] Figure 8 A flowchart illustrating a process (800) for adaptively controlling a heat treatment system according to an example embodiment of this disclosure is depicted. The process (800) can be implemented using a control system 600. For illustrative and discussion purposes, Figure 8 The steps are described in a specific order. Those skilled in the art using the disclosure provided herein will understand that the individual steps of any method described herein can be omitted, extended, performed concurrently, rearranged, and / or modified in various ways without departing from the scope of this disclosure. Furthermore, various additional steps (not shown) may be performed without departing from the scope of this disclosure.

[0092] In (802), the process may include: determining an estimated actual workpiece temperature based at least in part on data associated with the workpiece temperature of a first workpiece during workpiece heat treatment. For example, in Figure 6 In the illustrated embodiment, the actual workpiece temperature estimate 608 can be determined at least in part based on the system controller output 640 of the heat treatment system 606, which is associated with the workpiece temperature of the first workpiece of the heat treatment system 606.

[0093] In (804), the process may include: obtaining a simulated temperature estimate for a first workpiece using a system model, the system model providing the simulated temperature estimate based on one or more model parameters and one or more controller outputs. For example, in Figure 6 In the illustrated embodiment, the simulated workpiece temperature estimate 610 can be determined at least in part based on the clone system controller output 618 to system model 616, which is associated with the workpiece temperature of the first workpiece of heat treatment system 606.

[0094] In (806), the process may include: adjusting one or more controller parameters of the system controller based at least in part on the difference between the simulated temperature estimate obtained using a system model and the actual temperature estimate. For example, in Figure 6In the illustrated embodiment, the trusted control tuner 614 can generate and implement a system controller parameter adjustment 638 that is substantially similar to the clone controller parameter adjustment 630. The clone controller parameter adjustment 630 is generated at least in part based on the difference between the simulated workpiece temperature estimate 610 and the temperature setpoint 602.

[0095] At (808), the process may include: controlling one or more operating parameters of the heat treatment system, at least in part based on controller parameters, by a system controller to adjust the workpiece temperature of the second workpiece during heat treatment. For example, in Figure 6 In the illustrated embodiment, system controller 604 uses system controller output 640 to control the operating parameters of heat treatment 606. System controller output 640 can be applied to heat treatment system 606 for the subsequent processing of a second workpiece by heat treatment system 606.

[0096] Figure 9 A flowchart illustrating a process (900) for adaptively controlling a heat treatment system according to an example embodiment of this disclosure is depicted. The process (900) can be implemented using a control system 200. For illustrative and discussion purposes, Figure 9 The steps are described in a specific order. Those skilled in the art using the disclosure provided herein will understand that the individual steps of any method described herein can be omitted, extended, performed concurrently, rearranged, and / or modified in various ways without departing from the scope of this disclosure. Furthermore, various additional steps (not shown) may be performed without departing from the scope of this disclosure.

[0097] In (902), the process may include: obtaining an actual temperature output from one or more thermal sensors attached to the test workpiece. For example, in Figure 2 In one embodiment, the heat treatment system may include a test workpiece. The test workpiece may include one or more temperature sensors (e.g., one or more thermocouples) to measure the heat of the test workpiece. The thermocouples are accurate enough to measure heat to provide an actual temperature output 218.

[0098] In (904), the process may include: obtaining a simulated temperature estimate for the test workpiece using a system model. For example, in Figure 2 In the illustrated embodiment, system model 206 can generate a simulated temperature estimate 216 associated with the test workpiece of heat treatment system 204.

[0099] In (906), the process may include: receiving one or more parameter modifications from a system learning routine associated with the system model; the one or more parameter modifications being at least partially based on the difference between the simulated temperature estimate and the actual temperature output. For example, in Figure 2In the illustrated embodiment, system learning routine 210 may receive the difference between the simulated temperature estimate 216 and the actual system output 218 as input. System learning routine 210 may generate one or more parameter modifications 214 based at least in part on the difference between the simulated temperature estimate 216 and the actual system output 218.

[0100] In (908), the process may include: applying one or more parameter modifications to the system model. For example, in Figure 2 In the embodiment shown, the control system 200 can apply parameter modification 214 received from the system learning routine 210 to the system model 206, thereby modifying the parameters of the system model 206.

[0101] Although the subject matter has been described in detail with respect to specific example embodiments thereof, it should be understood that those skilled in the art, upon gaining an understanding of the foregoing, can readily produce changes, variations, and equivalents to these embodiments. Therefore, the scope of this disclosure is exemplary and not restrictive, and this disclosure does not exclude the inclusion of such modifications, variations, and / or additions to the subject matter, which will be apparent to those skilled in the art.

Claims

1. A method for controlling a thermal processing system, the method comprising: setting, by a control system of the thermal processing system, a temperature setpoint curve, wherein the temperature setpoint curve specifies a plurality of temperatures that can be applied to a workpiece; determining, by the control system of the thermal processing system, an actual workpiece temperature estimate associated with the workpiece; comparing, by the control system of the thermal processing system, the actual workpiece temperature to one of the plurality of temperatures in the temperature setpoint curve; determining, by the control system of the thermal processing system, a difference between the actual workpiece temperature estimate and the temperature setpoint curve; adjusting, by the control system of the thermal processing system, an operating parameter of the thermal processing system based at least in part on the difference between the actual workpiece temperature estimate and the temperature setpoint curve; tracking, by the control system of the thermal processing system, an actual workpiece temperature specified by the temperature setpoint curve relative to the actual workpiece temperature estimate; and accessing, by the control system of the thermal processing system, a system model, wherein the system model provides a simulated temperature estimate associated with the workpiece that simulates the actual workpiece temperature estimate associated with the workpiece, the system model comprising a machine learning model.

2. The method of claim 1, wherein, the system model comprises a machine learning neural network.

3. The method of claim 1, wherein, the system model is trained using a learning routine.

4. The method of claim 3, wherein, the learning routine comprises: obtaining actual temperature output from one or more thermal sensors attached to a test workpiece; obtaining a simulated temperature estimate for the test workpiece using the system model; and modifying one or more model parameters of the system model based on a difference between the actual temperature output and the simulated temperature estimate.

5. The method of claim 3, wherein, the learning routine comprises: obtaining actual workpiece characteristics of a test workpiece; obtaining simulated workpiece characteristics for the test workpiece using the system model; and modifying one or more parameters of the system model based on a difference between the actual workpiece characteristics and the simulated workpiece characteristics.

6. A thermal processing system, comprising: a process chamber; a workpiece support operable to support a workpiece during thermal processing in the process chamber; one or more thermal sources operable to heat the workpiece in the process chamber during thermal processing of the workpiece; one or more sensors configured to obtain data associated with workpiece temperature; and a control system configured to perform operations for controlling the thermal processing system, the operations for controlling the thermal processing system comprising: determining an actual workpiece temperature estimate associated with the workpiece; comparing the actual workpiece temperature to one of a plurality of temperatures in a temperature setpoint curve; determining a difference between the actual workpiece temperature estimate and the temperature setpoint curve; adjusting an operating parameter of the thermal processing system based at least in part on the difference between the actual workpiece temperature estimate and the temperature setpoint curve; tracking an actual workpiece temperature specified by the temperature setpoint curve relative to the actual workpiece temperature estimate; and ​ accessing a system model, wherein the system model provides a simulated temperature estimate associated with the workpiece, the simulated temperature estimate simulating the actual workpiece temperature estimate associated with the workpiece, the system model comprising a machine learning model.

7. The thermal processing system of claim 6, wherein, The control system includes a trusted control tuner configured to adjust one or more system controller parameters of a system controller.

8. The thermal processing system of claim 6, wherein, The control system includes a cloned system controller operable to provide an output indicative of a controller parameter to a system model.

9. The thermal processing system of claim 8, wherein, The cloned system controller includes one or more cloned controller parameters that are adjustable by a cloned control tuner.

10. The thermal processing system of claim 7, wherein, The cloned control tuner is operable to modify the one or more cloned controller parameters of the cloned system controller based on one or more control tuner adjustments learned using a tuning learning routine.

11. The thermal processing system of claim 9, wherein, The cloned control tuner is a machine learning model that associates errors with adjustments in controller parameters.

12. The thermal processing system of claim 10, wherein, The tuning learning routine determines control tuner adjustments based on simulated temperature estimates of a workpiece during heat treatment.

13. The thermal processing system of claim 6, wherein, The system model is trained using a learning routine.

Images