Molecular beam epitaxy substrate temperature field control method and system based on artificial intelligence

Abstract

The application relates to the technical field of thermal field modeling and control, and specifically provides a molecular beam epitaxy substrate temperature field control method and system based on artificial intelligence, which comprises the following steps: acquiring full-area temperature data of a substrate surface, heating power of each region of a heating cavity and environmental parameters; inputting the heating power and the environmental parameters into a temperature field model to output a temperature distribution matrix of the substrate surface and temperature gradient data; predicting temperature prediction data according to sequence data of the temperature distribution matrix, the heating power and the environmental parameters in a historical time period; and inputting a state vector formed by the temperature distribution matrix, the temperature gradient data, heating power data and the temperature prediction data into a deep reinforcement learning model to output a heating power adjustment amount of each region to control the temperature field. The application solves the problems of insufficient prediction accuracy and real-time performance of substrate temperature in the related art, considers the influence of environmental factors and heating power on the substrate temperature, and improves the temperature field control capability.

Description

Technical Field

[0001] This invention relates to the field of thermal field modeling and control technology, and in particular to a method and system for controlling the temperature field of molecular beam epitaxy substrates based on artificial intelligence. Background Technology

[0002] Molecular beam epitaxy (MBE) achieves atomic-level precision thin film growth by precisely projecting molecular beams of various elements onto a heated substrate surface in an ultra-high vacuum environment. The precision of substrate temperature control directly determines the crystal quality, electrical properties, and optical properties of the epitaxial film. However, due to thermal inertia, adjustments are only made after deviations occur, resulting in a long response delay and a lack of ability to predict temperature anomalies.

[0003] In related technologies, the substrate temperature control system of molecular beam epitaxy equipment uses a combination of thermocouple sensors and proportional-integral-derivative (PID) controllers to control the temperature by adjusting the output of a DC resistance heater. On the one hand, relying solely on single-point or localized temperature measurement by thermocouples makes it difficult to reflect the spatial temperature distribution of large-size substrates, and it does not fully consider the nonlinear coupling relationship between environmental factors such as vacuum level and ambient temperature and heating power, making it unable to accurately capture the comprehensive impact of multiple variables on substrate temperature. On the other hand, traditional feedback control lacks predictive capabilities, requiring adjustment to be initiated only after a temperature deviation occurs. Due to thermal inertia, the response lag is significant, making it impossible to predict temperature anomalies caused by environmental disturbances or process adjustments, and also difficult to optimize temperature prediction accuracy using historical data.

[0004] There is currently no effective solution to the problem that the prediction accuracy of substrate temperature in related technologies is insufficient, and that the influence of environmental factors and heating power on substrate temperature is not considered, resulting in insufficient control of the temperature field. Summary of the Invention

[0005] The artificial intelligence-based molecular beam epitaxy substrate temperature field control method and system provided by the present invention at least solves the problem of insufficient prediction accuracy and real-time performance of substrate temperature in related technologies, and takes into account the influence of environmental factors and heating power on substrate temperature, thereby improving the temperature field control capability.

[0006] According to one aspect of the present invention, an artificial intelligence-based method for controlling the temperature field of a molecular beam epitaxy substrate is provided, comprising the following steps: acquiring full-area temperature data of the substrate surface, heating power data and environmental parameter data of each region of the heating cavity in which the substrate is located; inputting the heating power data and the environmental parameter data into a temperature field model, and outputting a temperature distribution matrix and temperature gradient data of the substrate surface from the temperature field model; wherein, the temperature field model is an optimized model that adaptively corrects the material parameters of a three-dimensional unsteady-state heat transfer partial differential equation using the full-area temperature data over a set time period; predicting temperature prediction data for a future time period based on the temperature distribution matrix, the heating power and the environmental parameters in the sequence data of a historical time period; inputting a state vector composed of the temperature distribution matrix, the temperature gradient data, the heating power data and the temperature prediction data into a trained deep reinforcement learning model, and outputting an adjustment amount of the heating power for each region from the deep reinforcement learning model to control the temperature field; wherein, the deep reinforcement learning model is obtained by training a deep reinforcement learning neural network model based on historical state vector data, heating power adjustment data and reward data during state changes.

[0007] As an optional approach, obtaining full-area temperature data of the substrate surface includes the following steps: acquiring a temperature distribution image of the substrate surface using an infrared thermal imager; acquiring temperature data at multiple locations of the heated molybdenum support of the substrate using multiple temperature sensors; converting the temperature distribution image into an infrared temperature distribution matrix, and weighted and fused with the corresponding temperature data to obtain full-area temperature data of the substrate surface.

[0008] As an optional approach, converting the temperature distribution image into an infrared temperature distribution matrix includes the following steps: geometrically correcting the pixel coordinates of the corresponding positions in the temperature distribution image data based on the actual physical coordinates of the substrate surface to obtain corrected image data; obtaining the temperature sensing point coordinates of the temperature sensor on the substrate based on the actual coordinates of the temperature sensor, the size of the heating molybdenum support, and the assembly relationship between the heating molybdenum support and the substrate; matching the temperature sensing point coordinates with the pixel coordinates to establish a mapping relationship between the grayscale values ​​in the image data and the temperature data, and calibrating the temperature data at the corresponding positions in the image data; dividing the image data with calibrated temperature data into multiple grids, calculating the temperature of each grid using bilinear interpolation, and summing them to obtain an infrared temperature distribution matrix.

[0009] As an optional approach, before inputting the heating power data and environmental parameter data into the temperature field model, the method further includes: dividing the substrate into multiple grid cells; establishing a three-dimensional unsteady-state heat transfer partial differential equation for each grid cell based on the fundamental equations of heat transfer, and setting boundary conditions; the boundary conditions include: heat flow boundary conditions on the bottom surface of the substrate and radiation boundary conditions on the top / side surfaces of the substrate; combining the various three-dimensional unsteady-state heat transfer partial differential equations to obtain a preliminary temperature field model of the substrate; and adaptively correcting the material parameters of the three-dimensional unsteady-state heat transfer partial differential equations using the full-area temperature data over a set time period to obtain an optimized temperature field model; wherein, the material parameters are a parameter vector obtained by combining density, specific heat capacity, and thermal conductivity.

[0010] As an optional approach, the material parameters of the three-dimensional unsteady heat transfer partial differential equation are adaptively corrected using the full-area temperature data over a set time period. This includes the following steps: when the difference between the predicted temperature data output by the preliminary temperature field model and the actual temperature data exceeds a set threshold, the full-area temperature data over the set time period is input into the parameter iteration equation. The parameter iteration equation states that the material parameter at the next time step is equal to the difference between the material parameter at the current time step and the momentum term at the next time step. The momentum term at the next time step is equal to the sum of the momentum term at the current time step multiplied by the momentum coefficient and the gradient term of the objective function with respect to the material parameter. The objective function is constructed based on the difference between the temperature prediction value output by the preliminary temperature field model and the actual temperature data. The corresponding material parameters are solved using the parameter iteration equation to adaptively correct the material parameters of the three-dimensional unsteady heat transfer partial differential equation.

[0011] As an optional approach, the expression for the parameter iteration equation is: ; ; ; ;in, A parameter vector representing the number of iterations t+1; A parameter vector representing the number of iterations t; The momentum term represents the iteration number t+1; The momentum term represents the number of iterations t; Indicates the momentum coefficient; Indicates the learning rate; This represents the objective function of the parameter vector at iteration number t; This represents the gradient term of the objective function with respect to the material parameters at iteration number t; Indicates the number of samples; This represents the predicted temperature of the i-th temperature measurement point at iteration number t; This represents the actual temperature of the i-th temperature measurement point; This indicates the number of temperature measurement points.

[0012] As an optional approach, predicting temperature forecast data for a future time period based on the temperature distribution matrix, heating power, and environmental parameter sequence data over a historical time period includes the following steps: inputting the sequence data corresponding to the temperature distribution matrix, heating power data, and environmental parameter data for a set time period into a prediction model; wherein, the prediction model is obtained by training a neural network model using the temperature distribution matrix, heating power data, and environmental parameter data over a historical time period; and outputting temperature forecast data for the future time period from the prediction model; wherein, the temperature forecast data includes: predicted temperature value, predicted temperature change rate value, and predicted temperature acceleration value.

[0013] As an optional approach, before the prediction model outputs temperature prediction data for future time periods, the method further includes the following steps: training a neural network model using the temperature distribution matrix, heating power data, and environmental parameter data from historical time periods; calculating the total loss of the temperature prediction data output by the neural network model based on a loss function during each training iteration; wherein the loss function is obtained by weighted fusion of the temperature prediction loss function, the rate of change prediction loss function, and the acceleration prediction loss function; calculating the gradient of each parameter in the neural network model in reverse based on the total loss; combining the gradient and the learning rate to obtain a decay term, subtracting the parameter from the corresponding decay term to obtain the updated parameter; and stopping training until a termination condition is met, thus obtaining the prediction model.

[0014] As an optional approach, before inputting the state vector composed of the temperature distribution matrix, the temperature gradient data, the heating power data, and the temperature prediction data into the trained deep reinforcement learning model, the method further includes the following steps: using historical state vector data, heating power adjustment data, and reward data during state changes as a training dataset; wherein, the reward data is reward value data calculated by the reward function based on state changes; the reward function is obtained by weighted fusion of reward items based on temperature uniformity, control accuracy, response speed, and energy efficiency; using the training dataset, with the state vector data at each time step as input, the corresponding heating power adjustment data as output, and the state vector data at the next time step combined with the reward data as environmental feedback, the deep reinforcement learning model is trained; until the termination condition is met, training is stopped, and the trained deep reinforcement learning model is obtained.

[0015] According to another aspect of the present invention, an artificial intelligence-based molecular beam epitaxy substrate temperature field control system is provided, comprising: a sensing module for acquiring full-area temperature data of the substrate surface, heating power data and environmental parameter data of each region of the heating cavity in which the substrate is located; a first control module connected to the sensing module for inputting the heating power data and the environmental parameter data into a temperature field model, and outputting a temperature distribution matrix and temperature gradient data of the substrate surface from the temperature field model; wherein, the temperature field model is an optimized model that adaptively corrects the material parameters of a three-dimensional unsteady-state heat transfer partial differential equation using the full-area temperature data over a set time period; and a second control module connected to the first control module and the sensing module for adjusting the temperature distribution matrix, the heating power, and the environmental parameters according to the temperature distribution matrix, the heating power, and the environmental parameters. The system comprises: a sequence of historical time period data; temperature prediction data for future time periods; a third control module, connected to the second control module, the first control module, and the sensing module, which inputs a state vector composed of the temperature distribution matrix, the temperature gradient data, the heating power data, and the temperature prediction data into a trained deep reinforcement learning model, and outputs the adjustment amount of the heating power for each region from the deep reinforcement learning model; wherein the deep reinforcement learning model is trained on a deep reinforcement learning neural network model based on historical state vector data, heating power adjustment amount data, and reward data during state changes; and an execution module, connected to the third control module, which controls the heating power of the heaters in each region based on the adjustment amount of the heating power for each region, thereby controlling the temperature field.

[0016] According to another aspect of the present invention, an electronic device is also provided, comprising: a processor, and a memory storing a program, the program including instructions that, when executed by the processor, cause the processor to perform the artificial intelligence-based molecular beam epitaxy substrate temperature field control method described in any of the preceding claims.

[0017] Compared with the prior art, the above-described technical solution of the present invention has the following advantages:

[0018] The present invention provides an artificial intelligence-based molecular beam epitaxy substrate temperature field control method and system. This method no longer relies on single-point or localized thermocouple temperature measurement, but instead acquires temperature data across the entire substrate surface. Simultaneously, it collects heating power and environmental parameters, inputting a temperature field model adaptively corrected from the full-area temperature data. This model accurately fits the nonlinear coupling relationships between multiple parameters, improving the temperature uniformity problem of large-size substrates and overcoming the shortcomings of traditional control methods that ignore the combined effects of environmental factors and heating power. Secondly, it predicts future temperature data using historical temperature distribution matrices, heating power, and environmental parameter sequence data. Based on the state vector composed of the current temperature distribution matrix, temperature gradient data, heating power data, and future temperature prediction data, it inputs a trained deep reinforcement learning model to calculate the required heating power adjustment for substrate temperature control in real time. Therefore, it can adjust the substrate temperature field in real time, avoiding the lag of traditional control methods that adjust only when deviations occur. This solves the problems of insufficient prediction accuracy and real-time performance of substrate temperature in related technologies, improves temperature field control capabilities, and considers the influence of environmental factors and heating power on substrate temperature, thus enhancing temperature field control performance. Attached Figure Description

[0019] To more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention, and those skilled in the art can obtain other embodiments based on these drawings without creative effort.

[0020] Figure 1 This is a flowchart of an artificial intelligence-based molecular beam epitaxy substrate temperature field control method according to an embodiment of the present invention.

[0021] Figure 2 This is a schematic diagram of the structure of an artificial intelligence-based molecular beam epitaxy substrate temperature field control system, which is an embodiment of the present invention.

[0022] Figure 3 This is a schematic diagram of the structure of an artificial intelligence-based molecular beam epitaxy substrate temperature field control system, which is an embodiment of the present invention.

[0023] Figure 4 This is a schematic diagram of the structure of the electronic device created by this invention. Detailed Implementation

[0024] Embodiments of the present invention will now be described in more detail with reference to the accompanying drawings. While some embodiments of the present invention are shown in the drawings, it should be understood that the present invention can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of the present invention. It should be understood that the drawings and embodiments of the present invention are for illustrative purposes only and are not intended to limit the scope of protection of the present invention.

[0025] like Figure 1 As shown, according to one aspect of the present invention, a method for controlling the temperature field of a molecular beam epitaxy substrate based on artificial intelligence is provided, comprising the following steps S101 to S104.

[0026] Step S101: Obtain the temperature data of the entire surface of the substrate, the heating power data and environmental parameter data of each area of ​​the heating cavity where the substrate is located.

[0027] Step S102: Input the heating power data and environmental parameter data into the temperature field model, and output the temperature distribution matrix and temperature gradient data of the substrate surface from the temperature field model; wherein, the temperature field model is an optimized model that uses the full-area temperature data of a set time period to adaptively correct the material parameters of the three-dimensional unsteady heat transfer partial differential equation.

[0028] Step S103: Based on the temperature distribution matrix, heating power, and environmental parameters in the historical time series data, predict the temperature forecast data for the future time period.

[0029] Step S104 involves inputting a state vector composed of a temperature distribution matrix, temperature gradient data, heating power data, and temperature prediction data into a trained deep reinforcement learning model. The deep reinforcement learning model then outputs the adjustment amount of heating power for each region to control the temperature field. The deep reinforcement learning model is trained on a deep reinforcement learning neural network model based on historical state vector data, heating power adjustment data, and reward data during state changes.

[0030] In step S101, the full-dimensional input data required for temperature field control is acquired to provide data support for subsequent model calculations and decisions. The temperature data of the entire substrate surface must cover the entire substrate surface, rather than a single point or local area, to avoid uneven substrate temperature caused by local measurements. Full-area temperature data acquisition can be achieved using devices such as high-precision infrared thermometers and embedded thermocouple arrays. The data format is an N×M pixel two-dimensional temperature distribution image to capture temperature differences at different locations on the substrate.

[0031] When measuring heating power data, the heating cavity is usually heated independently in multiple areas, such as the central area, the edge area, and the annular area. It is necessary to collect the actual power consumption of each heating area. By monitoring the power consumption of each area, the system can understand the working status of the heater and provide feedback information on heating power for the control algorithm.

[0032] Environmental parameter data includes factors that can cause temperature field fluctuations, such as vacuum level inside the cavity, cavity wall temperature, and cooling water temperature. Specifically, vacuum level can affect thermal radiation efficiency, cavity wall temperature can affect environmental heat dissipation, and if cooling water temperature is present, it can affect heat transfer. Such data needs to be collected in real time to compensate for environmental disturbances.

[0033] Step S102 transforms the heating power and environmental parameters into substrate temperature distribution and gradient data through the optimized temperature field model, thus solving the parameter mismatch problem of the traditional fixed model. The temperature field model is based on the three-dimensional unsteady heat transfer partial differential equation, which describes the heat conduction, heat radiation, and heat convection processes of the substrate in the time and space dimensions.

[0034] Adaptive calibration utilizes the full-area temperature data collected in step S101 over a set time period to reverse-correct the material parameters in the model. By minimizing the error between the model's calculated temperature and the actual measured temperature, the material parameter values ​​are dynamically adjusted to ensure the model always closely matches the actual heat transfer scenario. Traditional models often use fixed material parameters, but in reality, these parameters may change with temperature and process stages, leading to significant deviations between the model's calculated results and the actual temperature. Adaptive calibration can further optimize the model.

[0035] The temperature distribution matrix on the substrate surface is in the same format as the temperature data collected in step S101, which intuitively reflects the temperature values ​​at each point on the substrate and is used to determine temperature uniformity, such as whether the temperature difference between the edge and the center exceeds the process threshold.

[0036] Temperature gradient data includes in-plane gradients along the substrate surface and longitudinal gradients perpendicular to the substrate, providing direction for subsequent power adjustment. The in-plane gradient affects lateral film uniformity, while the longitudinal gradient perpendicular to the substrate affects interfacial diffusion. If the gradient is too large, local heating power needs to be adjusted accordingly.

[0037] Step S103 uses sequence data from historical time periods to predict and intervene in the substrate temperature in advance, avoiding the problem of traditional control methods that require adjustment after detecting temperature deviations, which can easily lead to large response delays and loss of control in critical process stages.

[0038] Specifically, the temperature distribution matrix in historical time series data can reflect the temperature change trend, the heating power in historical time series data can reflect the response characteristics of the heating system, and the environmental parameters in historical time series data can reflect the change pattern of disturbance factors.

[0039] Temperature forecast data for future time periods can be obtained using time-series forecasting models. By learning the temporal correlations of historical time-series data, these models can output temperature forecast data for future time periods, such as temperature distribution matrices and temperature gradient data for the next 10 seconds and 30 seconds, providing a basis for subsequent power regulation.

[0040] The aforementioned time series prediction model can employ deep learning models such as LSTM (Long Short-Term Memory) and Transformer (a deep learning model based on self-attention), or traditional time series models such as ARIMA (Autoregressive Integrated Moving Average). This embodiment preferably uses LSTM, specifically optimized for the characteristics of temperature time series data, which can effectively capture the temporal correlation of series data over historical time periods.

[0041] Reward data is an indicator for measuring the quality of actions. The design principle is to guide the temperature field closer to the target state. For example, if the temperature uniformity improves, the gradient decreases, and the predicted temperature approaches the target value after adjustment, a positive reward is given; if the temperature deviation increases after adjustment and exceeds the process threshold, a negative reward is given.

[0042] By using historical state vectors, corresponding power adjustment amounts, and reward values ​​for the state changes in the adjusted temperature field, a deep reinforcement neural network model is trained. For example, under which historical state vector should a certain power adjustment amount be adopted, and the state change after adjustment should be closest to the target state to obtain the maximum reward. Through continuous iterative learning, a stable decision-making strategy is eventually formed.

[0043] The trained deep reinforcement neural network model can receive the current state vector in real time, quickly output the optimal heating power adjustment amount for each region, apply it to the heating unit, and at the same time re-collect and feed back the adjusted new temperature data, heating power data and environmental parameter data to the model through step S101, forming a closed loop of perception, calculation, prediction, decision-making and feedback.

[0044] The artificial intelligence-based molecular beam epitaxy substrate temperature field control method and system provided by the present invention no longer relies on single-point or local temperature measurement by thermocouples, but instead acquires temperature data of the entire substrate surface area, and simultaneously collects heating power and environmental parameters. The temperature field model, which is adaptively corrected by the full-area temperature data, is input to accurately fit the nonlinear coupling relationship between multiple parameters, improves the problem of temperature unevenness of large-size substrates, and makes up for the shortcomings of traditional control that ignore the combined effects of environmental factors and heating power.

[0045] Secondly, it predicts future temperature data by using historical temperature distribution matrix, heating power and environmental parameter sequence data. Based on the state vector composed of current temperature distribution matrix, temperature gradient data, heating power data and future temperature prediction data, it inputs the trained deep reinforcement learning model to calculate the adjustment amount of heating power required for current substrate temperature control in real time. Therefore, it can adjust the temperature of the substrate temperature field in real time, avoiding the lag of traditional control that adjusts only when deviation occurs.

[0046] In summary, this embodiment solves the problems of insufficient prediction accuracy and real-time performance of substrate temperature in related technologies, improves temperature field uniformity, and significantly reduces the probability of temperature control anomalies caused by unexpected factors such as external interference and thermocouple temperature jumps.

[0047] As an optional approach, obtaining the full-area temperature data of the substrate surface in step S101 includes the following steps: acquiring a temperature distribution image of the substrate surface using an infrared thermal imager; acquiring temperature data at multiple locations of the heated molybdenum support of the substrate using multiple temperature sensors; converting the temperature distribution image into an infrared temperature distribution matrix and weighting and fusing it with the corresponding temperature data to obtain the full-area temperature data of the substrate surface.

[0048] An infrared thermal imager is a non-contact temperature measurement device. It receives infrared radiation energy emitted from a substrate surface, converts it into a corresponding temperature value, and ultimately outputs a temperature distribution image. The aforementioned infrared thermal imager can select either a mercury cadmium telluride (HCdT) or indium antimonide (IST) detector to ensure high sensitivity and a wide temperature measurement range.

[0049] The aforementioned infrared thermal imager must be matched to the substrate size. Choosing a thermal imager with a resolution pixel count corresponding to the substrate size ensures the ability to distinguish millimeter-level temperature differences on the substrate surface, avoiding localized temperature averaging due to insufficient pixel count. The specific substrate temperature in molecular beam epitaxy depends on the thin film material; therefore, an instrument with a temperature measurement range covering this interval and an accuracy of ±1℃ must be selected to reduce fundamental measurement errors. During heating, the molybdenum support temperature is not static; therefore, the frame rate of the infrared thermal imager must be higher than the time scale of temperature changes to ensure no dynamic temperature information is missed and to achieve real-time data acquisition.

[0050] For example, in this embodiment, the diameter of the heated molybdenum tray is about 475mm. The infrared thermal imager is selected with a resolution of 320*240 pixels and a temperature measurement accuracy of ±1℃ to ensure that the value meets the process deviation threshold. The frame rate reaches 50 frames per second, and the thermal imager acquires a complete temperature distribution image every 20ms. This time interval is much smaller than the typical temperature change time of the molybdenum tray (500ms-2s), which can completely record the dynamic response process of the temperature.

[0051] Temperature sensors directly contact key points of the heated molybdenum base to collect precise local temperature data, which serves as the true value calibration anchor point for the infrared temperature distribution matrix. The temperature sensors can be thermocouples or platinum resistance sensors; this embodiment preferably uses high-precision thermocouples. Nine temperature measurement points are distributed across the entire heated molybdenum base using a nine-point array strategy to comprehensively detect the temperature field distribution of the molybdenum base. The temperature measurement accuracy of each thermocouple reaches 0.1℃.

[0052] The infrared temperature distribution matrix provides spatial continuity across the entire region, while the sensor at the heated molybdenum support provides local temperature data. By assigning weights and weighting the data, the local deviations of the infrared temperature distribution matrix are corrected, and finally, the temperature data of the entire substrate surface is obtained.

[0053] Specifically, the system synchronizes the collected data in time. For temperature data collected by thermocouples, the system uses a Kalman filter algorithm to remove measurement noise. Since different sensors have different sampling frequencies and response characteristics, the system uses an interpolation algorithm to unify the temperature data and infrared temperature distribution matrix to the same time reference.

[0054] After geometric correction and temperature calibration of the time-synchronized temperature distribution image, it is converted into an infrared temperature distribution matrix corresponding to the substrate surface through a bilinear interpolation algorithm.

[0055] The fusion of the infrared temperature distribution matrix and temperature data employs a combination of weighted averaging and Bayesian estimation. For temperature measurements taken by thermocouples and infrared thermal imagers at the same location, the system assigns weighting coefficients based on their respective measurement accuracy and reliability, and then calculates the fused full-area temperature data.

[0056] The specific fusion formula is as follows: the temperature data of the whole area is equal to the thermocouple weight multiplied by the thermocouple temperature plus the infrared weight multiplied by the infrared temperature, where the thermocouple weight plus the infrared weight equals one, and the weight coefficient is dynamically adjusted according to the historical measurement error statistics of the thermocouple sensor and the infrared thermal imager.

[0057] In summary, in molecular beam epitaxy (MBE) substrate temperature field control, the accuracy and spatial integrity of the full-area temperature data are core prerequisites for subsequent model calculations and intelligent decision-making. Infrared thermal imagers provide global distribution, while temperature sensors enable precise local measurements. This embodiment, through a weighted fusion scheme of infrared thermal imagers and multi-point temperature sensors, combines the advantages of both devices, ultimately obtaining full-area temperature data that balances spatial continuity and numerical accuracy.

[0058] As an optional approach, converting a temperature distribution image into an infrared temperature distribution matrix includes the following steps: Based on the actual physical coordinates of the substrate surface, geometrically correct the pixel coordinates of the corresponding positions in the temperature distribution image data to obtain corrected image data; based on the actual coordinates of the temperature sensor, the size of the heating molybdenum support, and the assembly relationship between the heating molybdenum support and the substrate, obtain the coordinates of the temperature sensing point on the substrate; match the temperature sensing point coordinates with the pixel coordinates to establish a mapping relationship between the grayscale values ​​in the image data and the temperature data, and calibrate the temperature data at the corresponding positions in the image data; divide the image data with calibrated temperature data into multiple grids, calculate the temperature of each grid using bilinear interpolation, and summarize to obtain an infrared temperature distribution matrix.

[0059] Geometric correction can eliminate image distortion caused by mounting angle and optical distortion, ensuring that the pixel position in the image corresponds precisely to the actual physical coordinates of the substrate surface.

[0060] Specifically, a physical coordinate system is first established with the center of the substrate surface as the origin, the radial direction as the r-axis, and the circumferential direction as the θ-angle. Markers with known physical coordinates are then placed on the substrate surface. The pixel coordinates of these markers are identified in the temperature distribution image. By solving a system of linear equations, a mapping relationship between the actual physical coordinates and the pixel coordinates is established. Using a tool library, this mapping relationship is applied to the entire temperature distribution image to obtain the corrected image data.

[0061] The temperature sensor is mounted on a heated molybdenum support, and its measurements must be mapped to specific locations on the substrate surface for temperature calibration. This process requires establishing a mapping relationship between the sensor coordinates on the heated molybdenum support and the actual physical coordinates of the substrate.

[0062] Specifically, with the center of the molybdenum support as the origin, the coordinates of each temperature sensor in the molybdenum support coordinate system are obtained using precision measuring tools. The assembly of the substrate and the molybdenum support typically involves concentricity and rotational deviations, which need to be determined through assembly relationship calibration. Through translational and rotational corrections, the coordinates of the temperature sensors on the molybdenum support are converted into the coordinates of the sensing point in the actual physical coordinate system of the substrate.

[0063] The grayscale values ​​of image data do not correspond directly and linearly with the actual temperature. It is necessary to use precise data from a temperature sensor for calibration to establish a quantitative relationship between grayscale values ​​and temperature values.

[0064] Specifically, based on image data, the coordinates of the temperature sensing point of the temperature sensor are matched one-to-one with the pixel coordinates in the image data. The gray values ​​of the matching pixel coordinates and the corresponding temperature values ​​of the sensing point coordinates are recorded to form multiple sets of sample data. According to the principle of infrared imaging, the relationship between gray value and temperature approximately satisfies an exponential function. The mapping function obtained by model fitting is applied to the entire image data, and the corresponding temperature value is marked at the position of each pixel in the image data.

[0065] By meshing and interpolation, a high-precision temperature distribution matrix can be generated by filling the entire region with known temperature points. Specifically, based on the spatial resolution required for temperature field control, the substrate surface is divided into a uniform mesh. For each mesh point, if it is not directly calibrated by the temperature sensor, bilinear interpolation is performed using the four surrounding pixels with known temperatures—that is, the closest pixel coordinates determined through geometric correction—to ensure the spatial continuity of the temperature distribution and avoid local jumps.

[0066] The bilinear interpolation method described above is an interpolation method that calculates the value of any unknown point on a two-dimensional plane using four known data points. Its core principle is to first perform linear interpolation along one direction, then linear interpolation along another direction, ultimately obtaining a smoothly transitioned result. In substrate temperature field data processing, the entire region's grid is filled with temperatures calibrated by a small number of known temperature sensors to ensure the spatial continuity of the temperature distribution.

[0067] An infrared temperature distribution matrix is ​​obtained by arranging the temperature values ​​of all grid points in physical coordinate order, with each element in the matrix corresponding to the temperature value at the substrate surface. This matrix retains the full-area coverage characteristic of the infrared image and improves temperature accuracy through temperature sensor calibration and bilinear interpolation calculation, fully meeting the requirement of obtaining full-area temperature data of the substrate surface in step S101.

[0068] As an optional approach, before inputting the heating power data and environmental parameter data into the temperature field model in step S102, the method further includes: dividing the substrate into multiple grid cells; establishing a three-dimensional unsteady-state heat transfer partial differential equation for each grid cell based on the fundamental equations of heat transfer, and setting boundary conditions; the boundary conditions include: heat flow boundary conditions on the bottom surface of the substrate and radiation boundary conditions on the top / side surfaces of the substrate; combining the various three-dimensional unsteady-state heat transfer partial differential equations to obtain a preliminary temperature field model of the substrate; and adaptively correcting the material parameters of the three-dimensional unsteady-state heat transfer partial differential equations using full-area temperature data over a set time period to obtain an optimized temperature field model; wherein, the material parameters are parameter vectors obtained by combining density, specific heat capacity, and thermal conductivity.

[0069] The temperature field model is a three-dimensional heat transfer mathematical model of the substrate established based on the finite element method. This model divides the substrate into several tiny tetrahedral grid units, and the temperature distribution within each grid unit is assumed to be linear.

[0070] The temperature change of each grid cell follows the law of conservation of energy and Fourier's law of heat conduction. Based on this, the three-dimensional unsteady-state heat transfer partial differential equation is derived, which is the dynamic equilibrium between the change in internal energy of the cell and the heat flow entering / leaving. For any grid cell, its three-dimensional unsteady-state heat transfer partial differential equation is: Where: ρ is density, c pHere, T represents specific heat capacity, t represents temperature, and t represents time. Indicates the thermal conductivity term; This represents the Hamiltonian operator; k is the thermal conductivity. This is the heat source term per unit volume.

[0071] The above boundary conditions are those that determine a unique solution to the partial differential equation. These boundary conditions are determined by the relationship between the heat transfer boundary and the surrounding environment, specifically including the heat flow boundary conditions on the bottom surface of the substrate and the radiation boundary conditions on the top / side surfaces of the substrate. The density ρ and specific heat capacity c mentioned above... p The parameter vector obtained by combining the thermal conductivity k is the material parameter.

[0072] Specifically, the bottom surface of the substrate acquires heat through physical contact with the molybdenum support, with heat transfer primarily via conduction. The heat flux boundary condition is that the heat flux density at the bottom surface equals the product of the substrate's thermal conductivity and the normal derivative of the bottom surface temperature. The top or side surface of the substrate is exposed to the high-vacuum environment of the molecular beam epitaxy cavity, with heat transfer primarily via radiation. The bidirectional radiation between the top or side surface of the substrate and the cavity wall must be considered. The radiation boundary condition is that the heat flux density at the top or side surface equals the net radiative heat flux from the top or side surface of the substrate to the cavity wall.

[0073] Material parameters are a parameter vector obtained by combining density, specific heat capacity, and thermal conductivity. The material parameters in the initial temperature field model are fixed values, but the material parameters in reality will change with temperature and process stage. Adaptive correction uses full-area temperature data over a set time period to reverse-correct the material parameters, achieving dynamic matching between the temperature field model and the actual scenario.

[0074] Specifically, the temperature data for the entire region over a set time period is aligned with the temperature calculated by the temperature field model using time and spatial coordinates, ensuring that the temperature data for each set time period corresponds to a unique temperature calculated by the temperature field model. The difference between the temperature prediction output by the preliminary temperature field model and the actual temperature data is calculated and compared with a set threshold. When the difference exceeds the set threshold, the material parameters of each grid cell are corrected to minimize the difference between the temperature prediction output by the preliminary temperature field model and the actual temperature data until the difference drops to the set threshold. The current material parameters are then considered the optimal corrected material parameters. The corrected material parameters are then substituted into the preliminary temperature field model, and the equations are solved again to obtain the optimized temperature field model.

[0075] As an optional approach, adaptive correction of the material parameters in the three-dimensional unsteady-state heat transfer partial differential equation is performed using full-area temperature data over a set time period. This includes the following steps: When the difference between the predicted temperature data output by the initial temperature field model and the actual temperature data exceeds a set threshold, the full-area temperature data for the set time period is input into the parameter iteration equation. The parameter iteration equation states that the material parameter at the next time step is equal to the difference between the material parameter at the current time step and the momentum term at the next time step. The momentum term at the next time step is equal to the sum of the momentum term at the current time step multiplied by the momentum coefficient and the gradient term of the objective function with respect to the material parameters. The objective function is constructed based on the difference between the temperature prediction value output by the initial temperature field model and the actual temperature data. The corresponding material parameters are solved using the parameter iteration equation, thus adaptively correcting the material parameters of the three-dimensional unsteady-state heat transfer partial differential equation.

[0076] Adaptive correction is not continuously running; instead, it determines whether to start by setting an error threshold, thus ensuring model accuracy while avoiding unnecessary computational costs.

[0077] The parameter iteration equation is the core algorithm for material parameter correction. It utilizes gradient descent, where the gradient of the objective function guides the direction of parameter optimization. Combining this with a momentum term accelerates convergence and prevents oscillations, enabling iterative updates of material parameters. The parameter iteration equation states that the material parameter at the next time step equals the difference between the material parameter at the current time step and the momentum term at the next time step. Here, the material parameter is a parameter vector obtained by combining density, specific heat capacity, and thermal conductivity.

[0078] Through the aforementioned adaptive correction process, the material parameters of the three-dimensional unsteady heat transfer partial differential equation can track the changes in actual heat transfer characteristics in real time, significantly improving the prediction accuracy of the temperature field model. The optimized model can more accurately output the substrate temperature distribution matrix and temperature gradient data, providing reliable physical model support for subsequent temperature prediction and deep reinforcement learning decision-making.

[0079] As an alternative approach, the expression for the parametric iteration equation is:

[0080] (1).

[0081] (2).

[0082] (3).

[0083] (4).

[0084] in, A parameter vector representing the number of iterations t+1; A parameter vector representing the number of iterations t; The momentum term represents the iteration number t+1; The momentum term represents the number of iterations t; Indicates the momentum coefficient; Indicates the learning rate; This represents the objective function of the parameter vector at iteration number t; This represents the gradient term of the objective function with respect to the material parameters at iteration number t; Indicates the number of samples; This represents the predicted temperature of the i-th temperature measurement point at iteration number t; This represents the actual temperature of the i-th temperature measurement point; This indicates the number of temperature measurement points.

[0085] In formula (1), when initializing the parameter iteration equation, The initial parameter vector was selected from typical values ​​in the material handbook; the momentum term was initialized. =0, meaning there is no initial historical inertia; momentum coefficient =0.9; learning rate =5e-4, the threshold set by the objective function corresponds to the process requirements that the prediction error must meet, for example, setting the threshold to be less than 1K. 2 The corresponding average temperature error is less than 1℃. Iteration stops when the objective function value is less than a set threshold, and the current material parameters are the optimized material parameters. If the set threshold is not reached, the gradient term is calculated.

[0086] In formula (2), the momentum term at the next moment simulates inertia in physics, retaining the update trend of the current moment and avoiding drastic fluctuations in the parameter value. The momentum coefficient is a quantity that controls the magnitude of inertia; the larger the momentum coefficient, the stronger the historical trend is retained.

[0087] In formula (3), the gradient term guides the direction of parameter optimization. When the gradient term is positive, the parameter needs to be decreased; when the gradient term is negative, the parameter needs to be increased.

[0088] In formula (4), the objective function is a metric that measures the error between the temperature calculated by the temperature field model and the temperature data of the entire region over a set time period. The smaller the objective function value, the better the parameters. Based on the temperature data of the entire region over a set time period, the objective function can be defined as the mean square error between the predicted temperature and the actual temperature, taking into account both the squared amplification of the error and mathematical differentiability. Predicted temperature It is the result of density ρ and specific heat capacity c at time t. p The parameter vector obtained by combining the thermal conductivity k The solution is obtained by substituting the equations into the three-dimensional unsteady heat transfer partial differential equation. This embodiment introduces a momentum term to accelerate convergence and reduce oscillations by utilizing historical update trends, such as reducing the number of iterations from 100 to 30. The momentum term accumulates historical gradient directions; if the parameter optimization direction remains consistent, the momentum term accelerates parameter updates and reduces the number of iterations. When the gradient direction changes abruptly, the inertia of the momentum term buffers the abrupt change, preventing drastic parameter oscillations.

[0089] The aforementioned parameter iteration equations quantify the prediction error using the objective function, guide the parameter adjustment direction using the gradient term, and balance the convergence speed and stability using the momentum term. This constitutes a complete closed loop of error feedback, gradient calculation, and parameter optimization, ultimately enabling the predicted temperature of the temperature field model to accurately match the actual measured temperature. The parameter iteration equations achieve adaptive correction of material parameters through a mathematical approach. The optimized parameter vectors significantly improve the model's simulation accuracy of the substrate heat transfer process, providing a reliable physical model foundation for subsequent temperature prediction and intelligent control decisions.

[0090] As an optional approach, the step S103, which predicts temperature forecast data for a future time period based on the temperature distribution matrix, heating power, and environmental parameters in the historical time period sequence data, includes the following steps: inputting the sequence data corresponding to the temperature distribution matrix, heating power data, and environmental parameter data for a set time period into the prediction model; wherein, the prediction model is obtained by training a neural network model using the temperature distribution matrix, heating power data, and environmental parameter data of the historical time period; and outputting the temperature forecast data for the future time period from the prediction model; wherein, the temperature forecast data includes: temperature forecast value, temperature change rate forecast value, and temperature acceleration forecast value.

[0091] The prediction model inputting sequence data requires time window design for the input layer. The prediction model employs a multi-resolution time window design to capture sequence data corresponding to temperature distribution matrices, heating power data, and environmental parameter data at different time scales. The multi-resolution time window design includes short, medium, and long time windows. The short time window covers the most recent 30 seconds of data with a sampling interval of 1 second; the medium time window covers the most recent 5 minutes of data with a sampling interval of 5 seconds; and the long time window covers the most recent 30 minutes of data with a sampling interval of 30 seconds.

[0092] The predictive model learns the time dependencies in historical time-series data through a neural network model, such as the hysteresis response of temperature after changes in heating power. The neural network model can employ either a long short-term memory network or a gated recurrent unit.

[0093] The prediction model in this embodiment is preferably trained using a long short-term memory network architecture. The hidden layer of the long short-term memory network adopts a two-layer structure. The first layer contains 64 memory units, which are mainly responsible for capturing the sequence data corresponding to the short-term temperature distribution matrix, heating power data and environmental parameter data. The second layer contains 32 memory units, which are mainly responsible for learning the long-term temperature change trend.

[0094] Each memory unit contains three gating mechanisms: a forget gate, an input gate, and an output gate. The forget gate determines what historical information is forgotten, the input gate determines what historical information is stored, and the input gate consists of two parts: a sigmoid layer (a network layer with the sigmoid function as its core) which determines which values ​​to update, and a tanh layer (a network layer with the tanh function as its core) which creates a vector of candidate values. The output gate determines what to output based on the currently stored information.

[0095] The prediction model uses the output layer of a Long Short-Term Memory (LSTM) network to predict temperature change trends at various measurement points over the next 5 to 30 seconds. To improve prediction accuracy, the output layer employs a multi-task learning design, simultaneously predicting temperature, rate of temperature change, and temperature acceleration. This design better captures the dynamic characteristics of temperature changes, improving prediction accuracy and stability.

[0096] The output layer of the aforementioned multi-task learning design improves overall prediction accuracy through information sharing between tasks, avoiding the error accumulation caused by neglecting the correlation of derivatives in single-task learning. For example, in temperature prediction data, temperature value, rate of change, and temperature acceleration do not exist in isolation. The rate of change of temperature is the first derivative of the temperature value with respect to time, and the temperature acceleration is the second derivative of the temperature value with respect to time. The three have a strong mathematical correlation and physical coupling.

[0097] As an optional approach, before the prediction model outputs temperature prediction data for future time periods, the method further includes the following steps: training the neural network model using the temperature distribution matrix, heating power data, and environmental parameter data for historical time periods; calculating the total loss of the temperature prediction data output by the neural network model based on the loss function during each training session; wherein the loss function is obtained by weighted fusion of the temperature prediction loss function, the rate of change prediction loss function, and the acceleration prediction loss function; calculating the gradient of each parameter in the neural network model in reverse based on the total loss; combining the gradient and the learning rate to obtain the decay term, subtracting the parameters from the corresponding decay term to obtain the updated parameters; stopping training until the termination condition is met, thus obtaining the prediction model.

[0098] The training process of the neural network model employs the backpropagation algorithm and the Adam optimizer. The backpropagation algorithm is responsible for calculating the gradient of the loss function with respect to the parameters of each layer, while the Adam optimizer efficiently updates the parameters based on these gradients. Together, they achieve iterative optimization of the neural network model from random initialization to accurate prediction.

[0099] The loss function is obtained by weighted fusion of the temperature prediction loss function, the rate of change prediction loss function, and the acceleration prediction loss function.

[0100] The temperature prediction loss function mentioned above can use the mean square error to calculate the deviation between the predicted temperature value and the actual temperature value; the rate of change prediction loss function can use the mean square error to calculate the deviation between the predicted rate of change of temperature and the actual rate of change; the acceleration prediction loss function also uses the mean square error to calculate and measure the deviation between the predicted temperature acceleration and the actual acceleration.

[0101] The total loss equals the temperature prediction loss multiplied by weight one, plus the rate of change prediction loss multiplied by weight two, plus the acceleration prediction loss multiplied by weight three. By adjusting these three weight coefficients, the degree of importance the neural network model places on different prediction tasks can be controlled. The training data comes from historical process records, typically amounting to tens of thousands of hours of continuous monitoring records.

[0102] Neural networks use the backpropagation algorithm to update parameters using the total loss. The core idea is to start from the loss, calculate the gradient of the parameters layer by layer, and then adjust the parameters in the opposite direction of the gradient to minimize the loss.

[0103] The gradients of the parameters mentioned above are the partial derivatives of the total loss with respect to the neural network parameters. They reflect the degree to which changes in the parameters affect the loss; a positive gradient indicates that increasing the parameters leads to an increase in loss, requiring a reduction in the parameters; a negative gradient indicates the opposite. The gradients can be obtained by backpropagating the gradients layer by layer using the chain rule to obtain the gradient values ​​of all parameters.

[0104] To accelerate convergence and improve stability, the Adam optimizer is used during training. This optimizer adds a momentum term to gradient descent to accumulate historical gradients and reduce oscillations. Simultaneously, the learning rate is increased, and the gradient and learning rate are combined to obtain a decay term, which is dynamically adjusted to update the parameters. The training of the neural network model's parameters is controlled by setting a termination condition to manage the number of training iterations.

[0105] Specifically, a total loss threshold is set. When the total loss after multiple training epochs is less than the threshold, it indicates that the neural network model has stably learned the data patterns, and further training will not provide significant improvement. Training can then be stopped, and a predictive model is obtained. A maximum number of training epochs is preset to prevent excessively long training times. Even if the loss has not fully converged, training should stop when the maximum number of epochs is reached to avoid wasting resources. If the validation set loss increases continuously for several epochs, it indicates that the model is beginning to overfit. At this point, training should be terminated, and the optimal parameters before overfitting should be saved. After training terminates, the model parameters with the best performance on the validation set should be saved to ensure that the model has good generalization ability on unseen data.

[0106] In summary, through the training process described above, the prediction model can extract temperature change patterns from historical sequence data and ultimately output reliable temperature prediction data for future time periods. This future temperature prediction data provides crucial information for subsequent deep reinforcement learning control decisions, enabling temperature field control to evolve from passive response to active prediction. This is a core technological support for achieving high-precision temperature control of molecular beam epitaxy substrates.

[0107] As an optional approach, before inputting the state vector composed of the temperature distribution matrix, temperature gradient data, heating power data, and temperature prediction data into the trained deep reinforcement learning model in step S104, the method further includes the following steps: using historical state vector data, heating power adjustment data, and reward data during state changes as a training dataset; wherein, the reward data is the reward value data calculated by the reward function based on state changes; the reward function is obtained by weighted fusion of reward terms based on temperature uniformity, control accuracy, response speed, and energy efficiency; using the training dataset, with the state vector data at each time step as input, the corresponding heating power adjustment data as output, and the state vector data at the next time step combined with the reward data as environmental feedback, the deep reinforcement learning model is trained; until the termination condition is met, training is stopped, and the trained deep reinforcement learning model is obtained.

[0108] The training dataset is organized using historical state vector data, heating power adjustment data, and reward data during state changes. The training dataset contains enough interaction experience over a long period of time, covering different process stages and environmental disturbance scenarios. During training, batches of data are randomly sampled from the training dataset to avoid overfitting of the model due to temporal correlation between samples. New interaction experience can be continuously added after the model is deployed to achieve online learning.

[0109] In the reward function, temperature uniformity is calculated from the temperature distribution matrix, which can encourage uniform temperature distribution on the substrate surface and reduce local temperature differences.

[0110] The control accuracy is calculated from the temperature distribution matrix and the target temperature, which can encourage the temperature to approach the target value.

[0111] The response speed is calculated from the temperature gradient data, which can encourage the temperature to converge to the target quickly and reduce regulation lag.

[0112] The energy efficiency reward memory heating power adjustment amount is calculated, which can encourage the reduction of unnecessary power consumption and avoid frequent and large power adjustments.

[0113] The reward function integrates reward items from four dimensions: temperature uniformity, control accuracy, response speed, and energy efficiency. Through weighted balancing of multi-objective optimization, it directly defines the criteria for the quality of actions, thereby determining whether the strategy learned by the model meets the temperature control objective.

[0114] The deep reinforcement learning model trained in this embodiment employs the deep deterministic policy gradient method. This method is suitable for control problems in continuous action spaces and can directly output continuous power adjustment values. The algorithm consists of two parts: an actor network and a critic network. The actor network is responsible for outputting the optimal heating power adjustment amount based on the current state vector data; the critic network is responsible for evaluating the value of the current state and action, providing guidance for the training of the actor network.

[0115] The aforementioned actor network employs a fully connected neural network structure, containing three hidden layers with 256, 128, and 64 neurons respectively. The input layer receives state vector data, and the output layer outputs heating power adjustment data. To ensure that the output heating power adjustment is within a reasonable range, the output layer uses a tanh activation function, and a scaling factor is used to map the output to the actual power adjustment range.

[0116] The aforementioned critic network also employs a fully connected neural network structure, but introduces a heating power adjustment variable in the second layer to fuse the state vector and the heating power adjustment variable action. This design allows for a better evaluation of the value of state-action pairs.

[0117] When the loss function of the commentator network (in the form of mean squared error, i.e., the square of the difference between the predicted value and the target value) is consistently below the set threshold, and the actual temperature control indicators of the policy in the validation scenario (such as temperature uniformity error ≤ 2℃, control accuracy ≤ 1℃) meet the target, and the average cumulative reward converges stably for multiple rounds (such as 50 rounds) with a fluctuation range < 5%), the iteration stops and the policy performance is evaluated; otherwise, the above steps are repeated to continue training the deep reinforcement learning model.

[0118] To improve the stability and convergence speed of learning, this embodiment also employs experience replay and target network techniques. Experience replay stores the reward function and next-state vector data in a replay buffer as environmental feedback. During training, batches of data are randomly sampled for network updates, breaking the temporal correlation between data. The target network is a delayed copy of the current network used to calculate the target value, avoiding drastic fluctuations in the target value during training.

[0119] Through the training process described above, the deep reinforcement learning model can learn control strategies adapted to complex temperature field dynamics from historical experience, ultimately achieving an end-to-end decision that outputs the optimal power adjustment amount from the input state vector. This model can not only cope with environmental disturbances and changes in material parameters, but also achieve intelligent balance among multiple objectives, providing high-precision, adaptive control capabilities for the temperature field of molecular beam epitaxy substrates.

[0120] According to another aspect of the present invention, an artificial intelligence-based molecular beam epitaxy substrate temperature field control system is also provided, such as... Figure 2 As shown, it includes: a sensing module for acquiring full-area temperature data of the substrate surface, heating power data and environmental parameter data of each area of ​​the heating cavity where the substrate is located; a first control module, connected to the sensing module, for inputting the heating power data and environmental parameter data into a temperature field model, and outputting the temperature distribution matrix and temperature gradient data of the substrate surface from the temperature field model; wherein, the temperature field model is an optimized model that adaptively corrects the material parameters of the three-dimensional unsteady heat transfer partial differential equation using full-area temperature data over a set time period; and a second control module, connected to the first control module and the sensing module, for predicting future time based on the temperature distribution matrix, heating power and environmental parameter sequence data over a historical time period. The system consists of a temperature prediction module and a third control module, connected to the second, first, and sensing modules. The third control module takes a state vector composed of a temperature distribution matrix, temperature gradient data, heating power data, and temperature prediction data as input to a trained deep reinforcement learning model. The deep reinforcement learning model outputs the adjustment amount of heating power for each region. The deep reinforcement learning model is trained using historical state vector data, heating power adjustment data, and reward data during state changes. The execution module, connected to the third control module, controls the heating power of the heaters in each region based on the adjustment amount of heating power for each region, thereby controlling the temperature field.

[0121] The sensing module integrates functions such as temperature measurement, infrared thermal imager, and power measurement. Temperature measurement uses multiple thermocouples mounted on the heated molybdenum support to directly detect the temperature. The infrared thermal imager receives the infrared radiation energy emitted by the substrate surface, converts it into the corresponding temperature value, and finally outputs a temperature distribution image. Power measurement uses Hall sensors to detect the instantaneous values ​​of current and voltage, and calculates the real-time power through a digital signal processor.

[0122] The first control module is connected to the sensing module. It uses data obtained from the sensing module to adaptively correct the temperature field model. This adaptive correction uses full-area temperature data to correct the material parameters in the model, minimizing the error between the calculated temperature and the actual measured temperature. It dynamically adjusts the material parameter values ​​to ensure the model always closely matches the actual heat transfer scenario. Traditional models often use fixed material parameters, but in reality, these parameters may change with temperature and process stages, leading to significant deviations between the model's calculations and actual temperatures. Adaptive correction further optimizes the model.

[0123] The temperature distribution matrix output by the temperature field model intuitively reflects the temperature values ​​at various points on the substrate, and is used to determine temperature uniformity, such as whether the temperature difference between the edge and the center exceeds the process threshold.

[0124] The temperature gradient data output from the temperature field model includes the in-plane gradient along the substrate surface and the longitudinal gradient perpendicular to the substrate, providing direction for subsequent power adjustment. The in-plane gradient affects the lateral film uniformity, while the longitudinal gradient perpendicular to the substrate affects interface diffusion. If the gradient is too large, the local heating power needs to be adjusted accordingly.

[0125] The second control module is connected to the first control module and the sensing module. It can predict the temperature forecast data for future time periods by using the sequence data of historical time periods obtained by the sensing module and the first control module. This enables early intervention in substrate temperature and avoids the problem of large response delay and loss of control in critical process stages caused by traditional control methods that only adjust after detecting temperature deviations.

[0126] The temperature distribution matrix in the historical time period can reflect the temperature change trend in the prediction process; the heating power in the historical time period can reflect the response characteristics of the heating system in the prediction process; and the environmental parameters in the historical time period can reflect the changing patterns of disturbance factors in the prediction process.

[0127] Temperature forecast data for future time periods can be obtained using time-series forecasting models. By learning the temporal correlations of historical time-series data, these models can output temperature forecast data for future time periods, such as temperature distribution matrices and temperature gradient data for the next 10 seconds and 30 seconds, providing a basis for subsequent power regulation.

[0128] The third control module is connected to the second, first, and sensing modules. It acquires historical state vectors, corresponding power adjustment amounts, and reward values ​​for the adjusted temperature field state changes through these modules, and uses this information to train the deep reinforcement neural network model. For example, it determines which power adjustment amount should be used under which historical state vector to achieve the state change closest to the target state, thus maximizing the reward. Through continuous iterative learning, a stable decision-making strategy is ultimately formed.

[0129] The trained deep reinforcement neural network model can receive the current state vector in real time, quickly output the optimal heating power adjustment for each region, and apply it to the execution module. At the same time, the adjusted new temperature data, heating power data, and environmental parameter data are re-collected and fed back to the model through the perception module, forming a closed loop of perception, calculation, prediction, decision-making, and feedback.

[0130] The execution module is connected to the third module. Based on the heating power adjustment amount obtained from the third module, the execution module controls the multi-zone heating system through the intelligent power adjustment module to control the temperature field.

[0131] The multi-zone heating system divides the substrate stage into multiple independent heating zones, each equipped with a specially designed resistance heater. The heating system employs a three-zone heating structure, corresponding to the center, middle, and edge of the molybdenum support. Each heating unit uses a wound tantalum wire heater, or can be made with high-purity tungsten or molybdenum wire, exhibiting excellent high-temperature stability and uniform power distribution characteristics. The heater's power density is carefully designed to ensure uniform heat flow distribution at rated power. To improve heating efficiency and reduce heat loss, the heater is surrounded by high-temperature insulation material.

[0132] The intelligent power regulation module, based on pulse width modulation (PWM) principles, controls the average output power by adjusting the on-time ratio of the power switching transistors. The control accuracy can reach 0.1% of the rated power, with a response time of less than 0.1 seconds. Employing advanced power electronics technology, the intelligent power regulation module enables precise power control of each heating zone.

[0133] The execution module also features safety functions such as overcurrent protection and overtemperature protection. Overcurrent protection immediately cuts off power output when an abnormally high current is detected to prevent damage to the heater. Overtemperature protection monitors the temperature of the power regulation module itself and reduces or stops output power when the temperature exceeds a safe threshold. Overload protection parameters such as current and temperature can be set manually.

[0134] By implementing the above technical solution, the present invention can bring about the following significant technological advancements and economic benefits:

[0135] Regarding temperature uniformity, this invention accurately calculates the three-dimensional temperature distribution of the substrate using a temperature field model and optimizes the heating power distribution in each region using a multi-region collaborative control algorithm. This significantly reduces the temperature non-uniformity of large-size substrates from 5-8°C in traditional control systems to less than 2°C. This improvement directly enhances the thickness and composition uniformity of the epitaxial layer, reducing the performance difference of epitaxial wafers prepared in the same furnace on a large-size (475mm) heater to less than 5%, thereby greatly improving product yield and consistency.

[0136] Regarding system response speed, this invention employs a predictive model based on artificial intelligence, capable of predicting temperature change trends 5 to 30 seconds in advance and taking corresponding control actions. This predictive control mechanism reduces the system's temperature response time from the traditional tens of seconds to 3-5 seconds, improving the response speed by nearly ten times. This rapid response capability allows the system to better track rapidly changing temperature setpoint curves, meeting the stringent requirements of complex multilayer epitaxial processes.

[0137] In terms of control precision, this invention acquires more accurate and comprehensive temperature field information through multi-sensor information fusion technology. Combined with precise calculations of the temperature field model and intelligent predictions of the prediction model, the temperature control precision is improved from ±1.2℃ in traditional systems to ±0.5℃. This high-precision temperature control directly improves the crystal quality of the epitaxial layer, enhances the interface quality of the superlattice material, and ultimately improves the optoelectronic performance and reliability of the device.

[0138] In terms of predictive capability, the long short-term memory neural network of this invention possesses powerful time-series learning capabilities, achieving a prediction accuracy of over 95% after thorough training. The system can predict potential temperature anomalies in advance, such as temperature fluctuations caused by heater power attenuation, environmental disturbances, and changes in process parameters, and automatically take preventative measures. This predictive control significantly reduces the failure rate of epitaxial processes, improving production stability and predictability.

[0139] In terms of energy consumption optimization, this invention utilizes multi-objective optimization based on reward data to minimize energy consumption while ensuring temperature control accuracy. Through an intelligent power allocation strategy, the system avoids overheating and energy waste common in traditional control methods. In practical applications, the system can reduce total energy consumption by 5%-10% while maintaining the same control accuracy, thus reducing production costs, the heat dissipation burden on equipment, and environmental impact.

[0140] Regarding system adaptability, the adaptive correction learning mechanism of this invention enables the system to quickly adapt to different substrate materials, sizes, and process conditions. When changing to different types of substrates, the system can automatically adjust control parameters and strategies, reducing the adaptation time from the traditional 2-3 hours to 20-30 minutes. This strong adaptability greatly improves the flexibility of equipment use and production efficiency.

[0141] It should be noted that this embodiment also provides an optional implementation method, which will be described in detail below.

[0142] like Figure 3 As shown, the molecular beam epitaxy (MBE) equipment requires precise temperature control of a 4-inch gallium arsenide substrate on a 475mm diameter heater inside the MBE chamber. The process requires a substrate temperature of 580℃±2℃, a temperature uniformity of ±2.5℃ throughout the furnace, and a control accuracy of ±1℃.

[0143] When configuring the sensing module, a temperature sensor array is first installed on the existing substrate stage. Taking into account the size, shape, and heating temperature zone distribution characteristics of the molybdenum holder, the sensor array adopts a nine-point temperature measurement layout to fully cover the heating area.

[0144] The infrared thermal imager is installed at the bottom observation window of the MBE cavity. Its technical parameters are: pixel resolution 320x240, temperature measurement range 200-1000℃, temperature accuracy ±1℃, and frame rate 50 frames per second. It uses a high-infrared-transmittance sapphire window with a baffle above it, which is closed when not in use to prevent the window surface from being vaporized with epitaxial material.

[0145] The substrate heating system employs a three-zone heating structure, corresponding to the center, middle, and edge portions of the molybdenum support, with each support having its own heating unit. The three-zone heating structure of each heating unit in the substrate heating system is controlled by a three-zone DC power supply. Each heating unit uses a wound tantalum wire heater, and the power density ensures the substrate can be heated up to 1000°C.

[0146] The power monitoring module is equipped with an independent power measurement unit for each heating zone, using Hall effect sensors to detect current and voltage, and a digital signal processor to calculate the power value in real time. The measurement accuracy is 0.5% of full scale, and the data update frequency is 1000 times per second.

[0147] The molecular beam epitaxy (MBE) device is also connected to a computer, which includes a sensing module and a control module. The control module within the computer comprises a first control module, a second control module, and a third control module. The first control module first establishes a temperature field model of the 4-inch gallium arsenide substrate. The temperature field model uses the finite element method, dividing the substrate into approximately 50,000 tetrahedral elements, each with a characteristic dimension of approximately 2 mm. Boundary conditions are set to consider actual heat transfer: the bottom surface of the substrate is in contact with the tantalum wire heater, using a constant heat flux density boundary condition, calculated based on the power and contact area of ​​each heating region; the top and side surfaces of the substrate are in contact with the vacuum environment of the MBE cavity, using a radiative heat transfer boundary condition, with the radiative heat transfer coefficient calculated based on the substrate surface emissivity, ambient temperature, source furnace temperature, and radiation amount; the contact portion between the substrate and the support structure uses a convective heat transfer boundary condition.

[0148] Material parameter calibration process for the temperature field model: The system collected temperature data from the first 50 process cycles, totaling approximately 250 hours of continuous monitoring records, as the training dataset for material parameter calibration. The calibration algorithm used the mean square error between the measured temperatures at nine temperature measurement points and the model-calculated temperatures as the optimization target, and continuously adjusted key parameters in the model, such as thermal conductivity and boundary heat transfer coefficient, using the gradient descent method. After calibration, the model achieved a temperature prediction accuracy of ±0.8℃.

[0149] The second control module employs a long short-term memory neural network architecture. The network's input layer is designed with 21 nodes, including 9 historical temperature sequences, 3 historical power sequences, and the rest being historical environmental parameter sequences. Each historical sequence is 30 time steps long, corresponding to data from the past 30 seconds, with a data sampling interval of 1 second.

[0150] The network's hidden layers employ a two-layer long short-term memory (LSTM) structure. The first layer contains 64 memory cells, primarily used to capture short-term temperature variation characteristics and periodic patterns. The second layer contains 32 memory cells, primarily used to learn long-term temperature variation trends and substrate thermal inertia properties. Each memory cell contains an input gate, a forget gate, and an output gate, as well as a cell state and a hidden state.

[0151] The network's output layer has 9 nodes, corresponding to the predicted temperature values ​​for the next 15 seconds at 9 temperature measurement points. The output layer uses a linear activation function to directly output the absolute value of the temperature. To improve the stability of the prediction, the output also includes the prediction confidence interval, with a confidence level set to 95%.

[0152] The network was trained using the backpropagation algorithm and the Adam optimizer. Training data was derived from historical process records, totaling 500 hours of continuous data. Data preprocessing included normalization, denoising, and outlier removal. The loss function was the mean squared error, calculated as the sum of squares of the differences between the predicted and actual temperatures. The learning rate was set to 0.001, the batch size to 32, and the number of training epochs to 1000. To prevent overfitting, early stopping and dropout regularization techniques were employed.

[0153] The trained neural network performed as follows on the test dataset: the mean absolute error of temperature prediction was 0.4℃, and the prediction accuracy (the proportion of errors less than 1℃) reached 96.5%. For predictions within different time windows, the prediction accuracy was highest after 5 seconds, with an error of approximately 0.2℃; the prediction error after 15 seconds was approximately 0.5℃; and the prediction error after 30 seconds was approximately 0.8℃.

[0154] The third control module employs a deep deterministic policy gradient reinforcement learning algorithm. The state space is designed as a 15-dimensional vector, including the temperature values ​​of the current nine temperature measurement points, the power values ​​of the three heating zones, and the temperature gradient information calculated from the digital twin model. The action space is a three-dimensional continuous vector, representing the power adjustment amount of the three heating zones, with an adjustment range of ±100W.

[0155] The reward function comprehensively considers multiple control objectives, specifically in the form that the total reward equals the negative weighted sum of errors, where the weight for temperature uniformity error is 0.4, the weight for control accuracy error is 0.3, the weight for response speed error is 0.2, and the weight for energy efficiency error is 0.1. Temperature uniformity error is defined as the standard deviation of the temperatures at the nine measurement points; control accuracy error is defined as the root mean square of the deviations between the temperatures at each measurement point and the setpoint; response speed error is defined as the difference between the rate of temperature change and the expected rate of temperature change; and energy efficiency error is defined as the power consumed per unit temperature change.

[0156] Both the actor network and the critic network in the reinforcement learning algorithm employ a three-layer fully connected structure. The actor network has 256, 128, and 64 hidden layer nodes, respectively, and uses ReLU (Modified Linear Unit) activation function. The output layer uses the tanh activation function to limit the output range. The critic network has a similar structure, but introduces action vectors in the second hidden layer to achieve state-action value evaluation.

[0157] The algorithm employs an empirical replay technique for training, with a replay buffer size of 100,000 samples. Each training iteration randomly samples 64 samples from the buffer for network updates. To improve training stability, a soft update method is used to update the target network, with an update coefficient of 0.005. A noise strategy is also introduced during training to promote exploration; the noise is generated using the Ornstein-Uhlenbeck process.

[0158] After 100,000 training iterations, the reinforcement learning algorithm converged to a stable control strategy. Simulation results show that this strategy achieves temperature uniformity within ±1.2℃, control accuracy of ±0.3℃, system response time reduced to 8 seconds, and energy consumption decreased by 5%-10% compared to traditional control methods.

[0159] The execution module connects to the existing molecular beam epitaxy (MBE) equipment control system via an industrial Ethernet interface. The control module outputs adjustment parameters to the execution module, which then adjusts the DC power supply for the three temperature zones, ultimately achieving heating control of the MBE equipment control system. The interface protocol adopts the Modbus TCP (Industrial Communication Control Protocol) standard, supporting real-time data transmission and remote control. The system software adopts a modular design, including a sensing module, a control module, and a human-machine interface module.

[0160] The human-machine interface uses a touchscreen display to show real-time information such as the temperature field distribution map, temperature curves at each measuring point, predicted result curves, and control parameter status. Operators can use the interface to set temperature curves, select control modes, and adjust control parameters. The interface also provides functions such as historical data query, alarm record viewing, and system diagnostics.

[0161] To verify the actual control effect of the system, a three-month continuous production test was conducted. During the test, a total of 120 epitaxial processes were performed, each lasting 2-4 hours. The test results are summarized below:

[0162] Regarding temperature uniformity, the average temperature difference across the nine measuring points in the traditional control system was 8.5℃, with a maximum difference of 12.5℃. After adopting the intelligent control system, the average temperature difference was reduced to 1.5℃, with a maximum difference of 2.1℃, demonstrating a significant improvement.

[0163] In terms of response speed, when the set temperature changes by 10°C, the traditional control system takes an average of 37 seconds to reach a new steady state, while the intelligent control system only takes 8 seconds, which is nearly 5 times faster.

[0164] In terms of control accuracy, the standard deviation of temperature control in traditional control systems is 1.2℃, while the standard deviation of intelligent control systems is reduced to 0.3℃, improving control accuracy by 4 times.

[0165] In terms of energy efficiency, under the same process conditions, the total energy consumption of the intelligent control system is reduced by 8% compared with the traditional control system, saving approximately RMB 100,000 in electricity costs annually.

[0166] In terms of epitaxial quality, the thickness uniformity of epitaxial wafers prepared using the intelligent control system has been improved from ±8% to ±3%, and the consistency of electrical properties has also been significantly improved. The device yield has increased from 63% to 78%.

[0167] According to another aspect of the invention, an electronic device is also provided, comprising: a processor, and a memory storing a program, the program including instructions that, when executed by the processor, cause the processor to perform any of the above-described artificial intelligence-based molecular beam epitaxy substrate temperature field control methods.

[0168] An embodiment of the present invention also provides an electronic device, including: at least one processor; and a memory communicatively connected to the at least one processor. The memory stores a computer program executable by the at least one processor, which, when executed by the at least one processor, causes the electronic device to perform the method of the embodiment of the present invention.

[0169] refer to Figure 4 This is a structural block diagram of an electronic device for a server or client, representing an embodiment of the present invention, and is an example of a hardware device that can be applied to various aspects of the present invention. The electronic device is intended to represent various forms of digital electronic computer devices, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device can also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the present invention described and / or claimed herein.

[0170] like Figure 4 As shown, the electronic device includes a computing unit 501, which can perform various appropriate actions and processes based on a computer program stored in a read-only memory (ROM) 502 or a computer program loaded from a storage unit 508 into a random access memory (RAM) 503. The RAM 503 may also store various programs and data required for the operation of the electronic device. The computing unit 501, ROM 502, and RAM 503 are interconnected via a bus 504. An input / output (I / O) interface 505 is also connected to the bus 504.

[0171] Multiple components in the electronic device are connected to I / O interface 505, including: input unit 506, output unit 507, storage unit 508, and communication unit 509. Input unit 506 can be any type of device capable of inputting information into the electronic device. Input unit 506 can receive input digital or character information and generate key signal inputs related to user settings and / or function control of the electronic device. Output unit 507 can be any type of device capable of presenting information and may include, but is not limited to, a display, speaker, video / audio output terminal, vibrator, and / or printer. Storage unit 508 may include, but is not limited to, disks and optical discs. Communication unit 509 allows the electronic device to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks, and may include, but is not limited to, modems, network cards, infrared communication devices, and / or wireless communication transceivers, such as Bluetooth devices, WiFi devices, WiMax devices, cellular communication devices, and / or the like.

[0172] The computing unit 501 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the computing unit 501 include, but are not limited to, CPUs, graphics processing units (GPUs), various special-purpose artificial intelligence (AI) computing units, various computing units running machine learning model algorithms, digital signal processors (DSPs), and any suitable processor, controller, microcontroller, etc. The computing unit 501 performs the various methods and processes described above. For example, in some embodiments, the method embodiments of the present invention can be implemented as computer programs tangibly contained in a machine-readable medium, such as storage unit 508. In some embodiments, part or all of the computer program can be loaded and / or installed on an electronic device via ROM 502 and / or communication unit 509. In some embodiments, the computing unit 501 can be configured to perform the methods described above by any other suitable means (e.g., by means of firmware). The computer program for implementing the methods of the embodiments of the present invention can be written in any combination of one or more programming languages. These computer programs can be provided to the processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing apparatus such that when executed by the processor or controller, the functions / operations specified in the flowcharts and / or block diagrams are implemented. Computer programs can be executed entirely on a machine, partially on a machine, or as standalone software packages, partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0173] In the context of embodiments of this invention, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable signal medium may include, but is not limited to, electronic, magnetic, optical, electromagnetic, or infrared systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0174] It should be noted that the term "comprising" and its variations used in the embodiments of this invention are open-ended, meaning "including but not limited to". The term "based on" means "at least partially based on". The term "one embodiment" means "at least one embodiment"; the term "another embodiment" means "at least one additional embodiment"; the term "some embodiments" means "at least some embodiments". The modifications of "one" and "a plurality" mentioned in the embodiments of this invention are illustrative and not restrictive, and those skilled in the art should understand that unless explicitly indicated otherwise in the context, they should be understood as "one or more".

[0175] The user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in the embodiments of this invention are all information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant countries and regions, and corresponding operation entry points are provided for users to choose to authorize or refuse.

[0176] The steps described in the method embodiments provided by the present invention can be performed in different orders and / or in parallel. Furthermore, the method embodiments may include additional steps and / or omit the steps shown. The scope of protection of the present invention is not limited in this respect.

[0177] The term "embodiment" in this specification refers to a specific feature, structure, or characteristic described in connection with an embodiment that may be included in at least one embodiment of the invention. The appearance of this phrase in various places throughout the specification does not necessarily imply the same embodiment, nor does it imply independence or alternativeity from other embodiments. The various embodiments in this specification are described in a related manner, with reference to each other for similar or identical parts. In particular, for apparatus, device, and system embodiments, since they are substantially similar to method embodiments, the description is relatively simple, and relevant details are referred to in the description of the method embodiments.

[0178] The above-described embodiments are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of protection. It should be noted that those skilled in the art can make various modifications and improvements without departing from the inventive concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the appended claims.

Claims

1. A method for controlling the temperature field of a molecular beam epitaxy substrate based on artificial intelligence, characterized in that, Includes the following steps: The temperature distribution image on the substrate surface was acquired using an infrared thermal imager; Temperature data at multiple locations on the heated molybdenum support of the substrate are collected using multiple temperature sensors. Based on the actual physical coordinates of the substrate surface, the pixel coordinates of the corresponding positions in the temperature distribution image data are geometrically corrected to obtain the corrected image data; Based on the actual coordinates of the temperature sensor, the dimensions of the heated molybdenum support, and the assembly relationship between the heated molybdenum support and the substrate, the coordinates of the temperature sensing point of the temperature sensor on the substrate are obtained. Match the coordinates of the temperature sensing point with the coordinates of the pixel to establish a mapping relationship between the grayscale value in the image data and the temperature data, and calibrate the temperature data at the corresponding position in the image data; The image data for calibrating temperature data is divided into multiple grids, and the temperature of each grid is calculated by bilinear interpolation. The results are then combined to obtain an infrared temperature distribution matrix. The infrared temperature distribution matrix is ​​weighted and fused with the corresponding temperature data to obtain the full-area temperature data of the substrate surface; Acquire heating power data and environmental parameter data for each region of the heating cavity containing the substrate; The heating power data and the environmental parameter data are input into the temperature field model, and the temperature field model outputs the temperature distribution matrix and temperature gradient data of the substrate surface; wherein, the temperature field model is an optimized model that adaptively corrects the material parameters of the three-dimensional unsteady heat transfer partial differential equation using the full-area temperature data of a set time period. Based on the temperature distribution matrix, the heating power, and the sequence data of the environmental parameters over a historical time period, predict the temperature forecast data for the future time period. The state vector, composed of the temperature distribution matrix, the temperature gradient data, the heating power data, and the temperature prediction data, is input into the trained deep reinforcement learning model. The deep reinforcement learning model outputs the adjustment amount of the heating power for each region to control the temperature field. The deep reinforcement learning model is trained on a deep reinforcement learning neural network model based on historical state vector data, heating power adjustment data, and reward data during state changes.

2. The method for controlling the temperature field of a molecular beam epitaxy substrate based on artificial intelligence according to claim 1, characterized in that, Before inputting the heating power data and the environmental parameter data into the temperature field model, the method further includes: The substrate is divided into multiple grid units; Based on the fundamental equations of heat transfer, a three-dimensional unsteady heat transfer partial differential equation is established for each grid cell, and boundary conditions are set; the boundary conditions include: the heat flow boundary conditions of the bottom surface of the substrate and the radiation boundary conditions of the top / side surface of the substrate. By combining the various three-dimensional unsteady heat transfer partial differential equations, a preliminary temperature field model of the substrate is obtained. The material parameters of the three-dimensional unsteady heat transfer partial differential equation are adaptively corrected using the full-region temperature data over a set time period to obtain an optimized temperature field model; wherein, the material parameters are a parameter vector obtained by combining density, specific heat capacity and thermal conductivity.

3. The method for controlling the temperature field of a molecular beam epitaxy substrate based on artificial intelligence according to claim 2, characterized in that, Adaptive correction of the material parameters of the three-dimensional unsteady heat transfer partial differential equation is performed using the full-region temperature data over a set time period, including the following steps: When the difference between the predicted temperature data output by the preliminary temperature field model and the actual temperature data exceeds a set threshold, the temperature data of the entire region for a set time period is input into the parameter iteration equation. The parameter iteration equation states that the material parameter at the next moment is equal to the difference between the material parameter at the current moment and the momentum term at the next moment. The momentum term at the next moment is equal to the sum of the momentum term at the current moment multiplied by the momentum coefficient and the gradient term of the objective function with respect to the material parameter. The objective function is constructed based on the difference between the temperature prediction value output by the preliminary temperature field model and the actual temperature data. The material parameters of the three-dimensional unsteady heat transfer partial differential equation are adaptively corrected by solving the corresponding material parameters through the parameter iteration equation.

4. The method for controlling the temperature field of a molecular beam epitaxy substrate based on artificial intelligence according to claim 3, characterized in that, The expression for the parameter iteration equation is: ； ； ； ； in, A parameter vector representing the number of iterations t+1; A parameter vector representing the number of iterations t; The momentum term represents the iteration number t+1; The momentum term represents the number of iterations t; Indicates the momentum coefficient; Indicates the learning rate; This represents the objective function of the parameter vector at iteration number t; This represents the gradient term of the objective function with respect to the material parameters at iteration number t; Indicates the number of samples; This represents the predicted temperature of the i-th temperature measurement point at iteration number t; This represents the actual temperature of the i-th temperature measurement point; This indicates the number of temperature measurement points.

5. The method for controlling the temperature field of a molecular beam epitaxy substrate based on artificial intelligence according to claim 1, characterized in that, Based on the temperature distribution matrix, the heating power, and the historical time series data of the environmental parameters, the temperature prediction data for future time periods is generated, including the following steps: The sequence data corresponding to the temperature distribution matrix, heating power data, and environmental parameter data for a set time period are input into the prediction model; wherein, the prediction model is obtained by training a neural network model using the temperature distribution matrix, heating power data, and environmental parameter data for historical time periods; The prediction model outputs temperature prediction data for a future time period; wherein, the temperature prediction data includes: temperature prediction value, temperature change rate prediction value, and temperature acceleration prediction value.

6. The method for controlling the temperature field of a molecular beam epitaxy substrate based on artificial intelligence according to claim 5, characterized in that, Before the prediction model outputs temperature forecast data for future time periods, the method further includes the following steps: The neural network model is trained using the temperature distribution matrix, heating power data, and environmental parameter data from historical time periods. During each training session, the total loss of the temperature prediction data output by the neural network model is calculated based on the loss function; wherein, the loss function is obtained by weighted fusion of the temperature prediction loss function, the rate of change prediction loss function, and the acceleration prediction loss function. Based on the total loss, the gradients of each parameter in the neural network model are calculated in reverse. The gradient and learning rate are combined to obtain a decay term. The parameters are subtracted from the corresponding decay term to obtain the updated parameters. Training continues until the termination condition is met, at which point the prediction model is obtained.

7. The method for controlling the temperature field of a molecular beam epitaxy substrate based on artificial intelligence according to claim 1, characterized in that, Before inputting the state vector, composed of the temperature distribution matrix, the temperature gradient data, the heating power data, and the temperature prediction data, into the trained deep reinforcement learning model, the method further includes the following steps: Historical state vector data, heating power adjustment data, and reward data during state changes are used as the training dataset; wherein, the reward data is the reward value data calculated by the reward function based on the state change; the reward function is obtained by weighted fusion based on reward items such as temperature uniformity, control accuracy, response speed, and energy efficiency. Using the training dataset, the state vector data at each time step is used as input, the corresponding heating power adjustment data is used as output, and the state vector data at the next time step is combined with the reward data as environmental feedback to train a deep reinforcement learning model. Training continues until the termination condition is met, at which point the training stops and a well-trained deep reinforcement learning model is obtained.

8. A molecular beam epitaxy substrate temperature field control system based on artificial intelligence, characterized in that, include: The sensing module is used to acquire temperature distribution images of the substrate surface using an infrared thermal imager; Temperature data at multiple locations on the heated molybdenum support of the substrate are collected by multiple temperature sensors; based on the actual physical coordinates of the substrate surface, the pixel coordinates of the corresponding positions of the temperature distribution image data are geometrically corrected to obtain corrected image data; based on the actual coordinates of the temperature sensors, the size of the heated molybdenum support, and the assembly relationship between the heated molybdenum support and the substrate, the coordinates of the temperature sensing point of the temperature sensor on the substrate are obtained. Match the coordinates of the temperature sensing point with the coordinates of the pixel to establish a mapping relationship between the grayscale value in the image data and the temperature data, and calibrate the temperature data at the corresponding position in the image data; The image data for calibrating temperature data is divided into multiple grids, and the temperature of each grid is calculated by bilinear interpolation. The results are then aggregated to obtain an infrared temperature distribution matrix. The infrared temperature distribution matrix is ​​weighted and fused with the corresponding temperature data to obtain the full-area temperature data of the substrate surface. The heating power data and environmental parameter data of each area of ​​the heating cavity where the substrate is located are obtained. The first control module, connected to the sensing module, is used to input the heating power data and the environmental parameter data into the temperature field model, and output the temperature distribution matrix and temperature gradient data of the substrate surface from the temperature field model; wherein, the temperature field model is an optimized model that uses the full-area temperature data of a set time period to adaptively correct the material parameters of the three-dimensional unsteady heat transfer partial differential equation. The second control module, connected to the first control module and the sensing module, is used to predict temperature forecast data for a future time period based on the temperature distribution matrix, the heating power, and the sequence data of the environmental parameters in the historical time period. The third control module, connected to the second control module, the first control module, and the sensing module, is used to input a state vector composed of the temperature distribution matrix, the temperature gradient data, the heating power data, and the temperature prediction data into a trained deep reinforcement learning model, and the deep reinforcement learning model outputs the adjustment amount of the heating power for each region; wherein, the deep reinforcement learning model is trained on a deep reinforcement learning neural network model based on historical state vector data, heating power adjustment data, and reward data during state changes; The execution module, connected to the third control module, is used to control the heating power of the heaters in each of the regions based on the adjustment amount of the heating power in each region, thereby controlling the temperature field.

9. An electronic device, comprising: A processor and a memory storing a program, characterized in that the program includes instructions that, when executed by the processor, cause the processor to perform the method according to any one of claims 1 to 7.

Images