A method for temperature control of hot plate uniformity based on PINN

By using a closed-loop control method based on PINN, the problems of difficult multi-zone collaborative control of the hot plate and insufficient energy consumption optimization are solved. This method enables continuous temperature field reconstruction and uniform temperature control of the hot plate, thereby improving the system's adaptability and energy efficiency.

CN123114152APending Publication Date: 2026-11-17HANGZHOU XIANDAN THERMAL POWER TECHNOLOGY CO LTD
View PDF 1 Cites 0 Cited by

Patent Information

Application Number
CN202611266434.5
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-20
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing hot plates face difficulties in multi-zone collaborative control and lack energy consumption optimization capabilities, making it difficult to achieve uniform temperature control.

Method used

A closed-loop control method based on Physical Information Neural Network (PINN) is adopted. Through hot plate partitioning, parameter acquisition, transient heat conduction model construction, PINN model training and optimization, temperature field reconstruction and power optimization are achieved, and closed-loop control is performed.

Benefits of technology

It achieves continuous temperature field reconstruction of the hot plate, identifies hot and cold spots between sensors and in the edge area, reduces average temperature error and peak-valley temperature difference, improves the system's adaptability to environmental and load changes, and reduces energy consumption and repeated adjustments.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123114152A_ABST
    Figure CN123114152A_ABST
Patent Text Reader

Abstract

This invention discloses a PINN-based method for temperature control of hot plates to ensure uniform temperature distribution, relating to the field of temperature control in industrial heating equipment. It aims to address the technical problems of difficulty in multi-zone collaborative control of hot plates and insufficient energy consumption optimization capabilities in existing methods. The method includes: periodically collecting various parameters of the hot plate after partitioning it; establishing a transient heat conduction physical model of the hot plate based on these parameters; constructing the framework and physical constraints of the PINN model based on the parameters and the transient heat conduction physical model; offline training using training samples to obtain a PINN temperature field prediction model, followed by online updates; based on the PINN temperature field prediction model, inputting the current power of each zone, ambient temperature, and operating parameters to obtain the predicted temperature of the discrete grid nodes on the hot plate surface; performing uniformity analysis and power optimization; and correcting and adjusting the optimized power adjustment sequence to obtain the actual power adjustment amount distributed to each zone. This achieves a closed-loop temperature control, significantly improving temperature control accuracy.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of temperature control for industrial heating equipment, and specifically to a method for temperature control based on the uniformity of a PINN-based hot plate. Background Technology

[0002] Heating plates or heating disks are widely used in metal heating, semiconductor processing, chemical reactions, food drying, and laboratory instrument applications. A typical heating plate consists of a heat-conducting disk body, multiple independent or semi-independent heating zones, temperature sensors, insulation structures, and a controller. Due to factors such as heating element layout, material thermal conductivity, structural thickness, edge heat dissipation, load contact conditions, and ambient temperature variations, the surface of the heating plate is prone to center-edge temperature differences, localized hot spots, and dynamic overshoot during the heating phase.

[0003] Existing PID control, fuzzy control, or zone control based on a small number of measurement points typically treat each measurement point as an independent control object, making it difficult to describe the lateral thermal coupling and continuous spatial temperature field inside the heat plate. The main problems are as follows: 1) Incomplete temperature field information. A limited number of sensors can only obtain the temperature at discrete measurement points, and cannot directly obtain the temperature distribution between sensors and in the edge areas; 2) Insufficient model adaptability. Fixed parameters or empirical rules are difficult to adapt to different target temperatures, ambient temperatures, load masses, and contact heat transfer conditions simultaneously; 3) Multi-zone coordinated control is difficult. Increasing the power of one zone will affect adjacent zones through heat conduction, and independent PID controllers are prone to repeated adjustments, oscillations, or localized overheating; 4) Insufficient energy consumption optimization capability. Traditional methods usually take the achievement of the set temperature at the measuring point as the sole objective, lacking a comprehensive consideration of temperature uniformity error, total power, power change rate, and over-temperature risk.

[0004] Chinese patent CN119835809A discloses a semiconductor processing device and its multi-zone heating plate and multi-zone temperature control driving method. It connects heating elements in different heating zones of the multi-zone heating plate within the semiconductor processing device by setting a current-limiting element. Each heating element undergoes both active and passive heating operations, and is subjected to one active heating power and at least one passive heating power. By controlling the ratio of active to passive heating power applied to the same heating element, the sum of the active and passive heating powers is made equal to a predetermined heating power for that element, ultimately achieving independent temperature control for each heating element. However, it still suffers from difficulties in multi-zone coordinated control. Summary of the Invention

[0005] To address the technical challenges of multi-zone collaborative control of existing hot plate heat exchangers and insufficient energy consumption optimization capabilities, this invention proposes a PINN-based hot plate temperature uniformity control method. This method achieves closed-loop control of "temperature acquisition - temperature field reconstruction - temperature uniformity evaluation - power optimization - control execution - model update," significantly improving temperature control accuracy.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a method for temperature control of hot plate uniformity based on PINN, characterized by comprising the following steps: S1: After partitioning the hot disk, periodically collect various parameters of the hot disk and establish a transient heat conduction physical model of the hot disk based on the parameters; S2, construct the framework and physical constraints of the PINN model based on various parameters and the transient heat conduction physical model of the hot plate; S3, the PINN temperature field prediction model is obtained through offline training using training samples and then updated online; S4, based on the PINN temperature field prediction model, inputs the power of each partition, ambient temperature and operating parameters at the current moment to obtain the predicted temperature of the discrete grid nodes on the surface of the heat plate; and performs temperature uniformity analysis and temperature uniformity power optimization. S5 corrects and adjusts the optimized power regulation sequence to obtain the actual power regulation amount distributed to each partition.

[0007] This invention provides a PINN-based method for controlling the uniformity of a hot plate temperature. This method takes the hot plate spatial coordinates, time, environmental conditions, and power of each heating zone as inputs, and the hot plate temperature as the output. Training constraints are formed by the transient heat conduction control equation, initial conditions, boundary conditions, and sensor measurement errors. The uniformity index is then calculated based on the predicted temperature field, and the power command for each heating zone is determined through rolling optimization, achieving closed-loop control of "temperature acquisition - temperature field reconstruction - uniformity evaluation - power optimization - control execution - model update".

[0008] Furthermore, after partitioning the hot disk, various parameters of the hot disk are periodically collected, including: S11, the surface of the hot plate is divided into several heating zones according to the heating structure, each zone corresponds to a heating element and a power drive module, and the hot plate operating parameters are recorded; S12, a temperature acquisition device is installed on the hot plate to periodically collect data on time, coordinates and temperature of each sensor, actual power of each zone, target temperature and ambient temperature.

[0009] This technical solution involves partitioning the hot plate, arranging sensors, and acquiring data to construct a physical model of transient heat conduction in the hot plate.

[0010] Furthermore, the transient heat conduction physical model of the heat plate adopts a transient heat conduction control relationship in the form of energy conservation. The transient heat conduction control relationship describes the balance between the change in heat storage in a unit volume of heat plate material, the heat conduction along each spatial direction, the volume heat source or surface heat flow generated by the heating element, and the heat dissipated to the environment or load.

[0011] Furthermore, the PINN model employs a fully connected neural network or a multilayer perceptron network with residual connections. The input of the PINN model includes at least normalized spatial coordinates, time, and power of each heating zone. The output of the PINN model is the predicted temperature.

[0012] Further, step S3 includes: S31, use historical running data to form a dataset and divide it into training set, validation set and test set; S32 generates collocation points for the governing equations, boundary collocation points, and initial collocation points inside the hot plate, at the boundary, and at the initial time, respectively. S33, Construct a weighted loss function and optimize network parameters; S34 determines whether to trigger online incremental updates based on the prediction error and a preset threshold.

[0013] Furthermore, the temperature uniformity analysis includes: Calculate the area-weighted average temperature, highest temperature, lowest temperature, peak-valley temperature difference, temperature standard deviation, and the proportion of area exceeding the deviation; The location, area, peak deviation, and region of hot and cold spots are identified through connected component analysis. The influence matrix of regional power on temperature is formed by using automatic differentiation or power perturbation.

[0014] Furthermore, the temperature uniformity power optimization includes: Using the power of each zone in the next few control cycles as the decision variable, a multi-objective rolling optimization model is established that takes into account temperature approximation, temperature uniformity, low energy consumption, power smoothing and over-temperature prevention, and optimization constraints are set. Depending on the real-time requirements, either a gradient-based algorithm or a heuristic algorithm is selected. For each candidate power combination, PINN quickly predicts the temperature field and calculates the objective function evaluation value.

[0015] Further, step S5 includes: The initial adjustment direction is generated based on hot or cold spots. After the cross-influence is corrected by the influence matrix, candidate power vectors are constructed. Gradient or heuristic algorithms are used to find the optimal solution within the constraints. PINN forward prediction is performed on the candidate schemes, and the scheme with the smallest comprehensive objective is selected. The first cycle command is executed in a rolling manner. Power dead zone and minimum hold time are set, and the control mode is switched when a fault or over-temperature occurs.

[0016] Furthermore, it also includes step S6, which includes: After executing the power commands for each zone, the error between the predicted temperature and the measured temperature is calculated. Continue using PINN when the error is within the acceptable range; when the error exceeds the limit for several consecutive periods, update using the latest sliding window data. When the average temperature deviation, peak-valley temperature difference, and temperature standard deviation all meet the threshold for several consecutive cycles, the system enters the heat preservation mode.

[0017] Furthermore, gradient-based algorithms or heuristic algorithms can be selected based on real-time requirements, including: When the PINN model can be directly differentiated, gradient-based algorithms are preferred for solving the power variables; when the control variables are discrete gears or the objective function is not differentiable, heuristic algorithms are used.

[0018] The present invention can bring the following beneficial effects: It can reconstruct the continuous temperature field of the hot plate and identify hot and cold spots between sensors and in the edge region; Improve the physical consistency of the model under insufficient sampling conditions by using heat conduction control equations, boundary conditions, and initial condition constraints; By optimizing power across multiple zones, the average temperature error, peak-valley temperature difference, temperature standard deviation, and out-of-range area ratio are reduced simultaneously. By employing rolling optimization, online updates, and security degradation mechanisms, the system's adaptability to changes in environment, load, and contact conditions is improved. Limiting fluctuations in total power and ineffective power while meeting temperature control and uniformity requirements helps reduce overheating, repeated adjustments, and energy waste. Attached Figure Description

[0019] Figure 1 This is a logic flowchart of a PINN-based hot plate temperature uniformity control method according to the present invention.

[0020] Figure 2 This is a flowchart illustrating the logic of PINN offline training and online updating in this invention. Detailed Implementation

[0021] Example 1:

[0022] To address the technical challenges of complex multi-zone collaborative control and insufficient energy consumption optimization capabilities in existing hot-disk systems, this embodiment proposes a hot-disk temperature uniformity control method based on a Physics-Informed Neural Network (PINN). (Refer to...) Figure 1 This includes the following steps.

[0023] Step S1: After partitioning the hot plate, periodically collect various parameters of the hot plate and establish a transient heat conduction physical model of the hot plate based on the parameters.

[0024] The periodic collection of various parameters of the hot disk after partitioning the hot disk in step S1 mainly includes the following sub-steps.

[0025] Step S11: Divide the surface of the hot plate into several heating zones according to the heating structure. Each zone corresponds to a heating element and a power drive module. Record the operating parameters of the hot plate.

[0026] In one optional embodiment, the surface of the hot plate is divided into M independently adjustable heating zones according to the heating structure, each zone corresponding to a heating element and a power drive module; in addition, the length, width, thickness, material density, specific heat capacity, thermal conductivity, as well as the position, area and rated power of the heating element are recorded; M is an integer greater than or equal to 2.

[0027] Step S12: Install temperature acquisition devices on the hot plate to periodically collect data on time, coordinates and temperature of each sensor, actual power of each zone, target temperature and ambient temperature.

[0028] In one optional embodiment, the temperature acquisition device uses temperature sensors, and N temperature sensors are arranged on or inside the hot plate; the temperature sensors preferentially cover the center, edge, corner, partition boundary and expected hot spot area of ​​the hot plate; In an alternative embodiment, an infrared imaging device may also be configured to acquire surface temperature images.

[0029] In one optional embodiment, the system collects data at a fixed sampling period, including time t, coordinates and temperatures of each temperature sensor, actual power of each zone, target temperature, and ambient temperature. Optional data acquisition includes load temperature, cooling fan speed, or contact pressure. The collected data undergoes time synchronization, outlier removal, and low-pass filtering.

[0030] The transient heat conduction physical model of the heat plate adopts the transient heat conduction control relationship in the form of energy conservation. The transient heat conduction control relationship describes the balance between the change of heat storage in a unit volume of heat plate material, the heat conduction along each spatial direction, the volume heat source or surface heat flow generated by the heating element, and the heat dissipated to the environment or load.

[0031] Specifically, the above-mentioned physical model of transient heat conduction in a hot plate includes at least the following variables and parameters: Spatial variables x, y, z and time variable t can describe any position and any time within the hot plate; The temperature T(x,y,z,t) to be determined is the main output of the PINN model; Material density ρ, specific heat capacity cp, and thermal conductivity k; when the material properties change with temperature, ρ(T), cp(T), and k(T) are used. The heating source term q(x,y,z,t) or boundary heat flux q(x,y,t) is calculated from the power Pm(t) of each heating zone, the location of the heating area, and the effective heating area. Ambient temperature Ta, convective heat transfer coefficient h, surface emissivity ε, and surrounding radiation temperature can describe the convective and radiative heat dissipation of the upper surface, lower surface, and sidewalls of the heat plate. The contact heat transfer coefficient hc or contact thermal resistance Rc when the hot plate is in contact with the heated load, as well as the load-side temperature Tl, can describe the heat transfer at the contact interface.

[0032] Boundary conditions are set according to the actual structure of the heat plate: the area covered by the heating element is set as a given heat flux or an equivalent volume heat source; the exposed surface is set as a convective heat dissipation boundary, with radiative heat dissipation superimposed if necessary; the insulated surface is set as an adiabatic or low heat transfer coefficient boundary; and the surface in contact with the load is set as a contact heat transfer boundary. The initial condition uses the measured temperature field at the start of heating; when the initial measurement points are limited, the interpolation result is first used as the initial temperature field, and then corrected by the PINN model combined with physical constraints.

[0033] Step S2: Construct the framework and physical constraints of the PINN model based on various parameters and the transient heat conduction physical model of the hot plate.

[0034] The PINN model employs a fully connected neural network or a multilayer perceptron network with residual connections. The input to the PINN model includes at least normalized spatial coordinates, time, and power of each heating zone. The output of the PINN model is the predicted temperature.

[0035] In one optional embodiment, the INN model employs a fully connected neural network or a multilayer perceptron network with residual connections. The network input includes at least normalized spatial coordinates (x, y, z), time t, and the power P1 to PM of each heating zone; depending on the operating conditions, ambient temperature Ta, load status, and heat transfer parameters may also be added. The network output is the predicted temperature. Of course, heat flux density or temperature gradient can also be increased as an auxiliary output.

[0036] In one alternative embodiment, physical constraints mainly include governing equation constraints, boundary condition constraints, initial condition constraints, measurement data constraints, and continuity and rationality constraints.

[0037] Governing equation constraints: Randomly select points inside the heat plate, substitute the automatic differentiation results of the network output with respect to time and space coordinates into the transient heat conduction energy balance, calculate the imbalance between the heat storage term, the heat conduction term in three directions and the heat source term, and make the residual approach zero.

[0038] Boundary condition constraints: Select boundary points on various boundaries so that the normal heat conduction and heat flow predicted by the network satisfy the given heat flow, convective heat transfer, radiative heat transfer, adiabatic heat transfer, or contact heat transfer relationship respectively.

[0039] Initial condition constraints: Ensure that the temperature field output by the network at the initial moment is consistent with the measured or interpolated initial temperature field.

[0040] Measurement data constraints: Ensure that the predicted temperature at the sensor coordinates matches the actual temperature, and different weights can be set according to the sensor accuracy.

[0041] Continuity and rationality constraints: Constrain the continuity of temperature or heat flow in adjacent areas, and add penalties for temperature predictions that significantly exceed the allowable range of the equipment.

[0042] Step S3: The PINN temperature field prediction model is obtained through offline training using training samples and then updated online.

[0043] Step S3 includes the following sub-steps.

[0044] Step S31: Use historical running data to form a dataset and divide it into training set, validation set and test set.

[0045] Step S32: Generate the coordinate points of the governing equations, boundary coordinate points, and initial coordinate points inside the hot plate, at the boundary, and at the initial time, respectively.

[0046] Step S33: Construct a weighted loss function and optimize network parameters.

[0047] Step S34: Determine whether to trigger online incremental update based on the prediction error and the preset threshold.

[0048] For more details, see the reference. Figure 2 Step S3 mainly includes two stages: offline initial training and online incremental update.

[0049] Establish training samples: Use historical operating data or simulation data under different target temperatures, different power combinations in different zones, different ambient temperatures and different load conditions to form a dataset; divide the dataset into training set, validation set and test set according to time period or operating condition to avoid data from the same heating process appearing in the training set and test set at the same time.

[0050] Generating physical collocation points: Collocation points for the governing equations, boundary collocation points, and initial collocation points are generated inside the hot plate, at the boundary, and at the initial time, respectively. The collocation point density is increased for edges, corners, and boundary regions with large temperature gradients.

[0051] Construct a weighted loss function: The total loss consists of measurement data error, control equation residuals, boundary condition residuals, initial condition residuals, and an optional regularization term. The weights can be preset or adaptively adjusted based on the gradient magnitude of each loss term to avoid any single term dominating training.

[0052] Optimize network parameters: First, use the Adam optimizer for batch iteration to achieve rapid network convergence; once the validation loss decreases slowly, switch to the L-BFGS optimizer to further reduce physical residuals. During training, learning rate decay, early stopping, and saving the best model are employed.

[0053] Training termination criteria: Offline training is completed when the validation set error no longer decreases after several consecutive rounds, and the root mean square error of sensor temperature, maximum temperature error, and physical equation residual are all below a set threshold.

[0054] Online update: When the device is running, new measurement data within the most recent sliding time window is added to a small batch of samples, and a small number of iterations are performed on the last few layers or all parameters of the network; when the prediction error does not exceed the threshold, the original model is maintained, and when it exceeds the threshold, incremental training is triggered, thus balancing real-time performance and model adaptability.

[0055] Step S4: Based on the PINN temperature field prediction model, input the power of each partition, ambient temperature and operating parameters at the current moment to obtain the predicted temperature of the discrete grid nodes on the hot plate surface; and perform temperature uniformity analysis and temperature uniformity power optimization.

[0056] Temperature uniformity analysis, including: Calculate the area-weighted average temperature, highest temperature, lowest temperature, peak-valley temperature difference, temperature standard deviation, and the proportion of area exceeding the deviation; The location, area, peak deviation, and region of hot and cold spots are identified through connected component analysis. The influence matrix of regional power on temperature is formed by using automatic differentiation or power perturbation.

[0057] More specifically, upon the arrival of the control cycle, the current time, power of each zone, ambient temperature, and operating parameters are input into the PINN model to obtain the predicted temperature of the discrete grid nodes on the hot plate surface. The temperature uniformity analysis employs the following methods: Calculate the area-weighted average temperature Tavg to characterize the overall temperature level of the hot plate; By calculating the maximum temperature Tmax, the minimum temperature Tmin, and the peak-valley temperature difference ΔTpv=Tmax-Tmin, it can be determined whether the maximum temperature difference requirement is met. Calculating the temperature standard deviation σT or root mean square deviation can evaluate the overall temperature dispersion and avoid relying solely on a single highest and lowest point; Calculate the absolute deviation of each node relative to the target temperature Tset, and calculate the proportion of the area exceeding the allowable temperature difference zone; Connectivity analysis is used to identify continuous hot spots and cold spots: first, grid nodes with temperatures higher than Tset+δ or lower than Tset-δ are binarized, and then merged into hot spot or cold spot areas according to spatial adjacency to obtain their location, area, peak deviation and heating zone; Based on the automatic differentiation of PINN, the sensitivity of temperature to the power of each zone is obtained, or forward prediction is performed by slightly perturbing the power of each zone to form an influence matrix of "zone power - temperature of each zone", which can determine which heating zone should increase or decrease power.

[0058] Temperature uniformity power optimization, including: Using the power of each zone in the next few control cycles as the decision variable, a multi-objective rolling optimization model is established that takes into account temperature approximation, temperature uniformity, low energy consumption, power smoothing and over-temperature prevention, and optimization constraints are set. Depending on the real-time requirements, either a gradient-based algorithm or a heuristic algorithm is selected. For each candidate power combination, PINN quickly predicts the temperature field and calculates the objective function evaluation value.

[0059] Real-time requirements necessitate the selection of gradient-based algorithms or heuristic algorithms, including: When the PINN model can be directly differentiated, gradient-based algorithms are preferred for solving the power variables; when the control variables are discrete gears or the objective function is not differentiable, heuristic algorithms are used.

[0060] Temperature uniformity power optimization is mainly achieved by constructing a temperature uniformity power optimization model. Specifically, a rolling optimization model is established using the power of each heating zone within the next H control cycles as the decision variable. The optimization objectives include: ensuring the heating plate temperature approaches the target value, achieving uniform surface temperature distribution, minimizing total energy consumption, ensuring gradual power changes between adjacent control cycles, and avoiding local overheating.

[0061] As an optional embodiment, the optimization constraints include at least: the power of each partition is between 0 and the rated power; the power change in a single cycle does not exceed the allowable ramp rate of the driver; the total power does not exceed the upper limit of the power supply system; the predicted maximum temperature does not exceed the safe temperature; and, if necessary, the power difference between adjacent partitions is constrained to prevent new local temperature differences from forming due to power abrupt changes.

[0062] As an optional implementation, the optimization algorithm can be selected according to real-time requirements: when the PINN model is directly differentiable, gradient-based sequential quadratic programming, projective gradient method, or Adam algorithm is preferred to solve for the power variables; when the control variables are discrete or the objective function is not differentiable, particle swarm optimization or genetic algorithm is used. Each candidate power combination is used by PINN to quickly predict the future temperature field and calculates an evaluation value based on the objective function.

[0063] Step S5: Correct and adjust the optimized power regulation sequence to obtain the actual power regulation amount sent to each partition.

[0064] Step S5 includes: generating an initial adjustment direction based on hot or cold spots, constructing a candidate power vector after correcting cross-influences with an influence matrix, and using gradient or heuristic algorithms to find the optimal solution within constraints; performing PINN forward prediction on the candidate solutions, selecting the solution with the smallest comprehensive objective, and executing the first cycle command in a rolling manner; setting the power dead zone and minimum hold time, and switching the control mode when encountering a fault or over-temperature.

[0065] Specifically, the power adjustment process includes the following steps.

[0066] The initial adjustment direction is generated based on the current predicted temperature field and hot / cold spot partitioning results: increase power in the partitions of cold spots and decrease power in the partitions of hot spots; at the same time, the cross-influence of adjacent partitions is corrected according to the power influence matrix.

[0067] Candidate power vectors are generated with the current power as the center. The gradient algorithm iteratively updates the power of each partition based on the partial derivative of the objective function with respect to the power; while the particle swarm optimization or genetic algorithm searches within the allowed power range and penalizes schemes that violate the power upper limit, temperature upper limit, and rate of change constraints.

[0068] PINN forward prediction is performed on the candidate schemes, and the future temperature standard deviation, peak-valley temperature difference, average temperature deviation and energy consumption are compared to select the scheme with the minimum comprehensive objective function.

[0069] The rolling time-domain approach executes only the power command of the first control cycle in the optimal sequence. The temperature is re-acquired and optimized again in the next cycle to correct model errors and external disturbances.

[0070] To avoid frequent fluctuations, a power dead zone and a minimum hold-up time are set. When the power change in a certain zone is less than the set threshold, the original power is maintained; in the event of sensor failure, communication abnormality, or predicted temperature exceeding the safety threshold, the system switches to power limiting or traditional PID safety control mode.

[0071] Step S6 includes: after executing the power command for each zone, calculating the error between the predicted temperature and the measured temperature; if the error is within the allowable range, continuing to use PINN; when the error exceeds the limit for several consecutive cycles, updating using the latest sliding window data; and entering the heat preservation mode when the average temperature deviation, peak-valley temperature difference, and temperature standard deviation all meet the threshold for several consecutive cycles.

[0072] Specifically, after executing the power commands for each zone, the controller reads the measured temperature in the next sampling cycle and calculates the error between the predicted temperature and the measured temperature. When the error is within the allowable range, the current PINN continues to be used; when the error exceeds the limit for multiple consecutive cycles, the latest sliding window data is used for online updates, or boundary parameters such as convective heat transfer coefficient and contact thermal resistance are re-identified.

[0073] When the average temperature deviation, peak-valley temperature difference, and temperature standard deviation all meet the set threshold for several consecutive cycles, the system enters the heat preservation mode. In the heat preservation mode, the optimization frequency and power change amplitude are reduced to reduce energy consumption.

[0074] The optimized frequency and power variation range reduced in heat preservation mode should be preset according to the actual situation.

[0075] Building upon this, this embodiment also proposes a PINN-based hot plate temperature uniformity control system, including a temperature acquisition module, an operating condition acquisition module, a data preprocessing module, a PINN temperature field prediction module, a temperature uniformity evaluation module, a power optimization module, a heating drive module, a safety protection module, and a model update module. All of these modules can be deployed in industrial computers, embedded controllers, or edge computing devices.

[0076] For the temperature acquisition module, the input is a thermocouple, RTD or infrared image, and the main output and function is discrete temperature data and timestamp.

[0077] For the PINN prediction module, the inputs are coordinates, time, power, and operating conditions, and the main outputs and functions are continuous temperature field and temperature gradient.

[0078] For the temperature uniformity evaluation module, the inputs are the predicted temperature field and the target temperature, and the main outputs and functions are the average temperature, peak-valley temperature difference, standard deviation, and hot / cold spot areas.

[0079] For the power optimization module, the inputs are temperature indicators, influence matrix, and power constraints, and the main output and function is the optimal power command for each heating zone.

[0080] For the safety protection module, its inputs are the upper temperature limit, sensor status, and communication status; its main outputs and functions are power limiting, shutdown, or PID standby control commands.

[0081] The model update module takes new measurement data and prediction error as input; its main output and function is the updated network parameters or boundary parameters.

[0082] The technical solution of this embodiment can bring about the following technical effects.

[0083] (1) Fusion of physical mechanism and sparse temperature measurement data: Transient heat storage, three-dimensional heat conduction, zoned heat source, convection / radiation heat dissipation and contact heat transfer are used together as PINN constraints to obtain a continuous temperature field under limited measurement point conditions, reducing unreasonable predictions caused by relying solely on data fitting.

[0084] (2) Temperature field indicators and hot spot identification for uniform temperature control: not only control the temperature of a single measuring point, but also calculate the peak-valley temperature difference, temperature standard deviation, out-of-range area ratio and hot spot / cold point connection area based on the prediction of the whole field, so that the uniform temperature target can be quantified and located.

[0085] (3) Multi-zone collaborative optimization based on power sensitivity: The influence relationship between zone power and overall temperature is obtained through PINN. Multiple heating zones are collaboratively adjusted under constraints such as power upper limit, rate of change and safe temperature to reduce repeated adjustments caused by thermal coupling between adjacent zones.

[0086] (4) Combining offline training with online incremental updates: Initial training is completed using historical data and physical points, and incremental updates of sliding windows are triggered based on real-time prediction errors to enable the model to adapt to changes in ambient temperature, load and contact conditions.

[0087] (5) Rolling time-domain closed-loop control and safety degradation: The optimizer predicts multiple control cycles in the future, but only executes the first cycle instruction and recalculates each time; at the same time, dead zone, power ramp-up rate, over-temperature limit and backup control mode are set to improve the stability and safety of real-time control.

[0088] Example 2:

[0089] Based on Example 1, a multi-zone industrial hot plate temperature control scheme is proposed. Taking a rectangular aluminum alloy hot plate as an example, nine independent heating zones are set below the hot plate, and temperature sensors are arranged on the surface at the center, four corners, midpoints of the four sides, and boundaries of the zones. The controller collects temperature and actual power every 1 second. The PINN input includes three-dimensional coordinates, time, nine power channels, and ambient temperature, and the output is the hot plate temperature. Offline training data covers heating up, steady-state heat preservation, different target temperatures, and different load conditions.

[0090] During operation, PINN predicts temperature on a regular grid on the hot plate surface, calculating the area-average temperature, peak-to-valley temperature difference, and temperature standard deviation. When a cold spot forms in the upper right corner, the temperature evaluation module identifies its area, lowest temperature, and corresponding zone; the power optimization module simultaneously considers the power sensitivity of this zone and adjacent zones, increasing the power of the relevant zones while appropriately reducing the power of the central zone where hot spots have formed, provided that the power does not exceed the rated power and power increase rate. The optimization results are executed for only one control cycle; the next cycle recalculates based on the measured temperature until the average temperature error and peak-to-valley temperature difference consistently meet the requirements.

[0091] Example 3:

[0092] Intelligent hot plate control with load-dependent contact changes: The hot plate can heat workpieces of varying masses and contact areas. After the workpiece is placed, the contact heat transfer conditions on the surface of the hot plate change. The system uses the temperature response within an initial number of sampling periods to jointly correct the contact heat transfer coefficient or contact thermal resistance, and uses the load state as the PINN input. If the prediction error exceeds a threshold for three consecutive control periods, online incremental training is triggered. Local overheating penalties and total energy consumption terms are added to the power optimization objective, enabling the hot plate to quickly recover to a uniform temperature after load changes, while avoiding overheating in unloaded areas.

[0093] Example 4:

[0094] Infrared temperature field-assisted calibration: During the equipment commissioning phase, a high-density temperature image of the hot plate surface is obtained using an infrared thermal imager, and emissivity and temperature calibration are performed with the contact sensors. Infrared images can expand the data coverage for offline training; during formal equipment operation, only a small number of contact sensors are needed, and the full-field temperature is reconstructed by PINN. This reduces the need for long-term use of high-cost imaging equipment while maintaining high temperature field prediction accuracy.

Claims

1. A method for temperature control of hot plate uniformity based on PINN, characterized in that, Includes the following steps: S1: After partitioning the hot disk, periodically collect various parameters of the hot disk and establish a transient heat conduction physical model of the hot disk based on the parameters; S2, construct the framework and physical constraints of the PINN model based on various parameters and the transient heat conduction physical model of the hot plate; S3, the PINN temperature field prediction model is obtained through offline training using training samples and then updated online; S4, based on the PINN temperature field prediction model, inputs the power of each partition, ambient temperature and operating parameters at the current moment, and obtains the predicted temperature of the discrete grid nodes on the surface of the hot plate; And perform temperature uniformity analysis and temperature uniformity power optimization; S5 corrects and adjusts the optimized power regulation sequence to obtain the actual power regulation amount distributed to each partition.

2. The method for temperature uniformity control of a hot plate based on PINN according to claim 1, characterized in that, After partitioning the hot disk, various parameters of the hot disk are periodically collected, including: S11, the surface of the hot plate is divided into several heating zones according to the heating structure, each zone corresponds to a heating element and a power drive module, and the hot plate operating parameters are recorded; S12, a temperature acquisition device is installed on the hot plate to periodically collect data on time, coordinates and temperature of each sensor, actual power of each zone, target temperature and ambient temperature.

3. The method for temperature uniformity control of a hot plate based on PINN according to claim 2, characterized in that, The transient heat conduction physical model of the heat plate adopts the transient heat conduction control relationship in the form of energy conservation. The transient heat conduction control relationship describes the balance between the change of heat storage in a unit volume of heat plate material, the heat conduction along each spatial direction, the volume heat source or surface heat flow generated by the heating element, and the heat lost to the environment or load.

4. A method for temperature control of a hot plate based on PINN according to claim 1, 2, or 3, characterized in that, The PINN model employs a fully connected neural network or a multilayer perceptron network with residual connections. The input of the PINN model includes at least normalized spatial coordinates, time, and power of each heating zone. The output of the PINN model is the predicted temperature.

5. The method for temperature control of hot plate uniformity based on PINN according to claim 4, characterized in that, Step S3 includes: S31, use historical running data to form a dataset and divide it into training set, validation set and test set; S32 generates collocation points for the governing equations, boundary collocation points, and initial collocation points inside the hot plate, at the boundary, and at the initial time, respectively. S33, Construct a weighted loss function and optimize network parameters; S34 determines whether to trigger online incremental updates based on the prediction error and a preset threshold.

6. A method for temperature control of hot plate uniformity based on PINN according to claim 1 or 5, characterized in that, The temperature uniformity analysis includes: Calculate the area-weighted average temperature, highest temperature, lowest temperature, peak-valley temperature difference, temperature standard deviation, and the proportion of area exceeding the deviation; The location, area, peak deviation, and region of hot and cold spots are identified through connected component analysis. The influence matrix of regional power on temperature is formed by using automatic differentiation or power perturbation.

7. The method for temperature uniformity control of a hot plate based on PINN according to claim 6, characterized in that, The temperature uniformity power optimization includes: Using the power of each zone in the next few control cycles as the decision variable, a multi-objective rolling optimization model is established that takes into account temperature approximation, temperature uniformity, low energy consumption, power smoothing and over-temperature prevention, and optimization constraints are set. Depending on the real-time requirements, either a gradient-based algorithm or a heuristic algorithm is selected. For each candidate power combination, PINN quickly predicts the temperature field and calculates the objective function evaluation value.

8. A method for temperature control of hot plate uniformity based on PINN according to claim 1 or 7, characterized in that, Step S5 includes: The initial adjustment direction is generated based on hot or cold spots. After the cross-influence is corrected by the influence matrix, candidate power vectors are constructed. Gradient or heuristic algorithms are used to find the optimal solution within the constraints. PINN forward prediction is performed on the candidate schemes, and the scheme with the smallest comprehensive objective is selected. The first cycle command is executed in a rolling manner. Power dead zone and minimum hold time are set, and the control mode is switched when a fault or over-temperature occurs.

9. The method for temperature control of hot plate uniformity based on PINN according to claim 1, characterized in that, It also includes step S6, which includes: After executing the power commands for each zone, the error between the predicted temperature and the measured temperature is calculated. Continue using PINN when the error is within the acceptable range; when the error exceeds the limit for several consecutive periods, update using the latest sliding window data. When the average temperature deviation, peak-valley temperature difference, and temperature standard deviation all meet the threshold for several consecutive cycles, the system enters the heat preservation mode.

10. A method for temperature control of hot plate uniformity based on PINN according to claim 7, characterized in that, Choose gradient-based algorithms or heuristic algorithms based on real-time requirements, including: When the PINN model can be directly differentiated, gradient-based algorithms are preferred for solving the power variables; when the control variables are discrete gears or the objective function is not differentiable, heuristic algorithms are used.

Citation Information

Patent Citations

  • Semiconductor processing equipment and multi-zone heating plate and multi-zone temperature control driving method thereof

    CN119835809A