Semiconductor heater temperature control method based on adaptive PID

Abstract

The application relates to the technical field of semiconductor manufacturing, and discloses a semiconductor heater temperature control method based on an adaptive PID. The method comprises the following steps: collecting semiconductor heater multi-monitoring-point temperature data in real time, constructing a temperature field distribution matrix to completely present the heater temperature distribution; generating a set temperature track based on the matrix, implementing closed-loop adjustment on the heating power through an adaptive PID controller, and making the temperature stably adhere to the set track; in the closed-loop adjustment, dynamically reconstructing a PID parameter fuzzy rule base according to a temperature fluctuation trend, and strengthening the adaptive capacity of the controller; when an abnormal temperature area is detected, triggering a multi-level compensation mechanism to perform phase correction on the controller, and quickly processing the abnormality. The method can optimize the temperature control precision and stability, cope with temperature fluctuations and abnormalities, and meet the stringent requirements of semiconductor manufacturing on temperature control.

Description

Technical Field

[0001] This invention relates to the field of semiconductor manufacturing technology, specifically to a semiconductor heater temperature control method based on adaptive PID. Background Technology

[0002] In the semiconductor manufacturing field, the temperature control accuracy of heaters directly affects the performance and yield of semiconductor devices. As semiconductor devices develop towards miniaturization and high integration, the requirements for heater temperature control are becoming increasingly stringent. It is necessary not only to achieve precise maintenance of the target temperature, but also to cope with the balanced regulation of the temperature field in multiple regions.

[0003] Currently, temperature control of semiconductor heaters mostly employs traditional PID control methods. These methods typically set control parameters based on temperature data from a single monitoring point, making it difficult to comprehensively reflect the overall temperature field distribution of the heater. Due to issues such as heat conduction hysteresis and uneven local heat loss in semiconductor heaters, data from a single monitoring point cannot accurately capture temperature differences in different areas, easily leading to an unbalanced temperature field distribution. This, in turn, can cause localized overheating or underheating during semiconductor device processing, affecting the device's physical characteristics and electrical performance.

[0004] Traditional PID control parameters are mostly fixed values, which cannot be dynamically adjusted according to temperature fluctuations during heater operation. When semiconductor manufacturing processes change, ambient temperature changes, or the heater shows slight aging, the fixed PID parameters are difficult to adapt to these dynamic changes, which can easily lead to problems such as overshoot, oscillation, or response delay in temperature control. This not only reduces the accuracy of temperature control but may also affect the production efficiency and product quality of semiconductor devices.

[0005] Existing temperature control methods are insufficient for handling areas with abnormal temperatures. When localized temperature anomalies are detected, simple power compensation or shutdown for inspection are often used, lacking a multi-level, precise compensation mechanism. Simple power compensation may affect the temperature of adjacent areas, further disrupting the temperature field equilibrium; while shutdown for inspection interrupts the production process, increasing production costs and time. These problems restrict the further development of semiconductor heater temperature control technology and fail to meet the current semiconductor manufacturing requirements for high-precision, high-stability temperature control. Summary of the Invention

[0006] The purpose of this invention is to provide a temperature control method for semiconductor heaters based on adaptive PID, so as to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides a semiconductor heater temperature control method based on adaptive PID, the method comprising:

[0008] Real-time acquisition of temperature data from multiple monitoring points of a semiconductor heater to construct a temperature field distribution matrix;

[0009] A set temperature trajectory is generated based on the temperature field distribution matrix, and the heating power of the semiconductor heater is adjusted in a closed loop using an adaptive PID controller.

[0010] During the closed-loop control process, the fuzzy rule base of PID parameters is dynamically reconstructed based on the temperature fluctuation trend.

[0011] When an abnormal temperature region is detected, a multi-level compensation mechanism is triggered to perform phase correction on the adaptive PID controller.

[0012] Preferably, generating a set temperature trajectory based on the temperature field distribution matrix includes:

[0013] Analyze the axial and radial temperature gradients in the temperature field distribution matrix;

[0014] Retrieve the corresponding temperature distribution template based on the current process formula;

[0015] By integrating the axial temperature gradient, radial temperature gradient, and temperature distribution template, a set temperature trajectory with time dimension constraints is generated.

[0016] Preferably, the heating power of the semiconductor heater is closed-loop regulated using an adaptive PID controller, including:

[0017] Calculate the deviation between real-time temperature data and the set temperature trajectory, as well as the rate of change of the deviation.

[0018] Based on the deviation and the rate of change of deviation, the fuzzy rule base of PID parameters is queried, and the proportional coefficient correction, integral coefficient correction, and derivative coefficient correction are output.

[0019] Adjust the real-time PID control parameters according to the proportional coefficient correction, integral coefficient correction, and derivative coefficient correction.

[0020] Preferably, the fuzzy rule base for PID parameters is dynamically reconstructed based on temperature fluctuation trends, including:

[0021] Extract the spectral characteristics of temperature fluctuations over N consecutive sampling periods;

[0022] The frequency bands whose amplitude exceeds a threshold in the spectral characteristics of the temperature fluctuations are identified as the dominant perturbation frequency bands.

[0023] The fuzzy rule membership function of PID parameters is reconstructed with the goal of suppressing the dominant disturbance frequency band.

[0024] Preferably, when an abnormal temperature region is detected, the following is included:

[0025] Calculate the acceleration of temperature change at each monitoring point;

[0026] Time series cluster analysis was performed on the temperature change acceleration;

[0027] The region with the largest cluster center value is marked as a temperature anomaly region.

[0028] Preferably, a multi-level compensation mechanism is triggered to perform phase correction on the adaptive PID controller, including:

[0029] Locate the nearest compensation node based on the spatial coordinates of the temperature anomaly area;

[0030] Activate the feedforward compensation channel of the nearest compensation node;

[0031] A correction signal with the opposite phase to the dominant disturbance frequency band is injected into the feedforward compensation channel.

[0032] Preferably, the method further includes:

[0033] A multi-dimensional evaluation index system for the temperature control process is established, which includes steady-state error index, overshoot index, and settling time index.

[0034] The multidimensional evaluation index system is weighted and fused to generate a comprehensive control effectiveness score;

[0035] The comprehensive control performance score is combined with the current control parameters and stored in the historical case library.

[0036] Preferably, the method further includes:

[0037] When a change in process formulation is detected, the most similar historical control parameter combination is retrieved from the historical case database;

[0038] The historical control parameter combination is injected into the adaptive PID controller for preheating initialization;

[0039] The preheating parameters are finely adjusted based on the deviation between the real-time temperature data and the set temperature trajectory.

[0040] Preferably, the method further includes:

[0041] Collect energy consumption distribution data and heat dissipation status data of semiconductor heaters, construct a Pareto front optimization model for temperature control accuracy and energy consumption, and output PID parameters to constrain the boundary through the Pareto front optimization model.

[0042] Preferably, the method further includes:

[0043] Real-time monitoring of changes in the coefficient of thermal expansion of semiconductor heater materials;

[0044] When the change in the coefficient of thermal expansion exceeds the tolerance threshold, the thermal deformation compensation algorithm is activated, and the compensation amount output by the thermal deformation compensation algorithm is superimposed on the output of the adaptive PID controller.

[0045] Compared with the prior art, the beneficial effects of the present invention are:

[0046] By collecting temperature data from multiple monitoring points in real time and constructing a temperature field distribution matrix, the overall temperature distribution of the semiconductor heater can be comprehensively and accurately reflected, overcoming the limitation of traditional single-point monitoring data in failing to capture regional temperature differences. With the help of the temperature field distribution matrix, the temperature change characteristics of different regions can be clearly understood, providing a more precise basis for subsequent temperature control decisions. This allows temperature control to no longer rely on local data but to be based on the overall temperature field, effectively avoiding temperature field imbalance caused by missing local temperature information. This helps maintain the overall temperature uniformity of the heater, ensuring that all regions of the semiconductor device are in a suitable temperature environment during processing.

[0047] In the temperature control stage, a set temperature trajectory is generated based on the temperature field distribution matrix, and an adaptive PID controller is used for closed-loop regulation, achieving dynamic adaptability of temperature control. Unlike traditional fixed-parameter PID control, this method can adjust the heating power in real time according to the actual changes in the temperature field, ensuring that the temperature always operates stably around the set trajectory. This closed-loop regulation method can respond promptly to subtle temperature fluctuations, avoiding problems such as overshoot and oscillation caused by fixed parameters, improving the stability and accuracy of temperature control. This allows the heater to better adapt to the temperature requirements of different stages in semiconductor manufacturing processes, reducing the adverse effects of temperature fluctuations on device performance.

[0048] During closed-loop regulation, the fuzzy rule base of PID parameters is dynamically reconstructed based on temperature fluctuation trends, further enhancing the adaptive capability of the PID controller. Temperature fluctuation trends reflect key information such as the rate, direction, and magnitude of current temperature changes. Reconstructing the fuzzy rule base based on this information allows for more targeted and rational adjustment of PID parameters. When faced with process switching, environmental changes, or slight heater aging, the reconstructed fuzzy rule base provides the PID controller with control rules more suited to the current operating conditions, ensuring that PID parameters are always in an optimal state. This avoids control parameter mismatch problems caused by changes in operating conditions, extends the effective adaptation cycle of the PID controller, reduces the frequency of manual parameter adjustments, and lowers operational complexity and human error.

[0049] When an abnormal temperature region is detected, a multi-level compensation mechanism is triggered to perform phase correction on the adaptive PID controller, effectively improving the ability to handle temperature anomalies. This multi-level compensation mechanism can adopt different levels of compensation measures based on factors such as the degree of temperature deviation and the range of influence of the abnormal region, rather than the traditional simple power compensation or shutdown. For minor temperature anomalies, precise compensation can be achieved by fine-tuning the PID controller phase, avoiding interference with the temperature of adjacent regions. For more severe anomalies, a higher-level compensation strategy can be activated to quickly restore the temperature of the abnormal region to the normal range while ensuring overall temperature field balance. This multi-level compensation method avoids secondary temperature problems caused by simple compensation, eliminates the need for shutdown, ensures the continuity of the production process, reduces production interruptions and cost increases caused by temperature anomalies, and also reduces the impact of abnormal temperatures on the production quality and efficiency of semiconductor devices. Attached Figure Description

[0050] Figure 1 This is a schematic diagram illustrating the working principle of the semiconductor heater temperature control method based on adaptive PID described in this invention.

[0051] Figure 2 The flowchart for the closed-loop regulation of an adaptive PID controller;

[0052] Figure 3 A flowchart for dynamically reconstructing the fuzzy rule base for PID parameters;

[0053] Figure 4 A flowchart for constructing a multidimensional evaluation index and historical case library;

[0054] Figure 5 This is a flowchart for preheating initialization when the process formula is changed. Detailed Implementation

[0055] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0056] Please see Figure 1 This invention provides a method for temperature control of a semiconductor heater based on adaptive PID control, the method comprising:

[0057] Real-time temperature data from multiple monitoring points on a semiconductor heater is acquired to construct a temperature field distribution matrix. Specifically, multiple temperature sensor arrays are uniformly arranged on the surface of the semiconductor heater, with each sensor acquiring real-time temperature values ​​at fixed sampling intervals. These temperature values ​​are organized into a two-dimensional matrix, where rows represent monitoring points at different spatial coordinates, and columns represent time-series data. The matrix dimensions are dynamically adjusted according to the heater's geometry and process requirements to ensure complete spatial coverage. A set temperature trajectory is generated based on the temperature field distribution matrix, and an adaptive PID controller is used to perform closed-loop regulation of the semiconductor heater's heating power. This controller drives the heating element by adjusting the power output signal in real time. During closed-loop regulation, the system continuously analyzes temperature fluctuation trends and dynamically reconstructs the fuzzy rule base for PID parameters. This includes monitoring statistical changes in temperature data and updating the rule base based on historical fluctuation patterns. When an abnormal temperature region is detected, a multi-level compensation mechanism is triggered to perform phase correction on the adaptive PID controller. Anomaly detection is based on temperature gradient analysis, and the compensation mechanism involves activating the feedforward control channel to correct phase shifts. The entire implementation process ensures the stability and response speed of temperature control.

[0058] Example 1: See Figure 2 In the operation of generating a set temperature trajectory based on the temperature field distribution matrix, the system performs gradient analysis. This operation processes the spatial data characteristics of the temperature field distribution matrix in both axial and radial dimensions. The axial temperature gradient analysis process unfolds along the central axis of the semiconductor heater, and the system extracts the temperature value sequence of continuous monitoring points located on the same axis in the matrix. For each pair of adjacent monitoring points, the temperature difference is calculated and divided by the physical distance between the two points to obtain the local temperature change slope. The system traverses all axial monitoring point pairs, performs arithmetic mean processing on the obtained local slope data, and outputs the axial average temperature gradient value. This value characterizes the overall temperature change trend along the length of the heater. The radial temperature gradient analysis is performed perpendicular to the axis direction. The system selects monitoring point data located on the same radius ring from the matrix. Taking the heater center as the origin, the average temperature of each ring region is calculated layer by layer outward. The difference between the average values ​​of adjacent ring regions is divided by the radial distance to obtain the radial temperature gradient value sequence. The system uses the central difference algorithm to process discrete radial data points to ensure the smoothness of gradient calculation. All gradient calculation results are temporarily stored in a high-speed cache for subsequent fusion processing.

[0059] The process recipe management module retrieves the corresponding temperature distribution template from the database based on the current production batch identifier. This template is a predefined two-dimensional matrix structure, with rows corresponding to the spatial coordinates of monitoring points and columns containing time node information. The template data reflects the target temperature distribution pattern for a specific process stage, including but not limited to heating rate, steady-state temperature value, and spatial uniformity requirements. After the template is loaded, the system performs a trajectory fusion operation. The fusion process introduces a weighted superposition algorithm to spatially register the axial and radial temperature gradient data with the temperature distribution template. The axial gradient data is assigned a first weight coefficient, the radial gradient data a second weight coefficient, and the temperature distribution template a baseline weight coefficient. Each weight coefficient is dynamically adjusted according to the process stage: increasing the weight of gradient data during the rapid heating stage and increasing the template weight during the isothermal stage. A time dimension constraint function is added during the superposition process, which uses the process duration as the independent variable and outputs a time compensation factor. The fusion algorithm ultimately generates a set temperature trajectory matrix, where the row vectors represent the target temperature time series of each monitoring point, and the column vectors contain the target spatial distribution of each time slice. The trajectory data is indexed by timestamps, accurate to the millisecond level of the control cycle.

[0060] The closed-loop power regulation process initiates the operation of the adaptive PID controller. The controller receives two input signals in real time: the current actual temperature value collected from the temperature sensor array, and the target value at the corresponding time node in the previously generated set temperature trajectory. The system performs deviation calculation operations at fixed control cycles. The deviation is defined as the algebraic difference between the actual temperature value and the target value; the calculation process iterates through all monitoring points to generate a deviation vector. The deviation change rate is derived from historical deviation data. The system records the deviation vector of the previous control cycle, performs element-wise difference operations with the current deviation vector, and then divides by the control cycle duration to obtain the change rate matrix. This matrix reflects the instantaneous trend of temperature deviation at each monitoring point.

[0061] The fuzzy rule base query operation is based on the deviation and deviation change rate matrix. The system inputs continuous deviation and change rate values ​​into the fuzzification interface. The fuzzification interface is configured with a predefined set of membership functions: the deviation is divided into seven fuzzy sets: negative large, negative medium, negative small, zero, positive small, positive medium, and positive large; the deviation change rate is divided into five fuzzy sets: negative fast, negative slow, zero, positive slow, and positive fast. Each input value calculates its membership degree in each fuzzy set through the membership function, generating a fuzzy input vector. The fuzzy inference engine performs matching according to a pre-set rule table, which uses an IF-THEN structure to describe the input-output relationship. Each rule corresponds to a fuzzy set mapping of three output items: proportional coefficient correction, integral coefficient correction, and differential coefficient correction. Rule triggering uses the min-max inference method: for each rule, the minimum value of the input condition membership degree is taken as the rule strength; for each output item, the maximum strength of all related rules is taken as the output membership degree. The defuzzification process uses the centroid method to calculate the precise output value, ultimately obtaining the specific values ​​of the proportional coefficient correction, integral coefficient correction, and differential coefficient correction.

[0062] The PID parameter adjustment module receives three correction values. The system maintains a real-time PID control parameter set, including the proportional coefficient KP, integral coefficient KI, and derivative coefficient KD. The adjustment operation employs an incremental update strategy: new proportional coefficient = original proportional coefficient + proportional coefficient correction; new integral coefficient = original integral coefficient × integral coefficient correction; new derivative coefficient = original derivative coefficient + derivative coefficient correction. The updated parameter set is immediately applied to the controller output calculation. The output calculation uses a positional PID algorithm: first, the current deviation vector is spatially averaged to obtain the global temperature deviation; second, the deviation is integrated over time to accumulate the historical deviation sum; finally, the difference between the current cycle deviation and the previous cycle deviation is calculated. The control output is obtained by a weighted sum of the proportional, integral, and derivative terms, with weights corresponding to the KP, KI, and KD parameters, respectively. The final output signal is converted into a voltage or current command to drive the power actuator of the semiconductor heater. The power actuator contains a solid-state relay or IGBT module, which linearly adjusts the input power of the heating element according to the command. The entire closed-loop control process is repeated in each control cycle, forming a continuous feedback control loop. Temperature data acquisition, trajectory comparison, rule base query, parameter adjustment, and power output all strictly adhere to timing constraints to ensure the real-time response capability of the control system. Matrix operations employ a parallel processing architecture, with distributed computing nodes processing data subsets from different monitoring point areas and synchronizing data via a high-speed bus to meet the time requirements for large-scale data processing across multiple monitoring points.

[0063] Example 2: See Figure 3During closed-loop regulation, the system dynamically reconstructs the fuzzy rule base of PID parameters. This process begins with the extraction of temperature fluctuation spectral features. The system is configured with a fixed-length data buffer to store temperature data sequences for N consecutive sampling periods. The number of sampling periods N is set by system configuration parameters, and its value is determined based on the heater's thermal inertia time constant to ensure coverage of the complete temperature fluctuation cycle. Each sampling period contains temperature readings from all monitoring points, forming a space-time matrix. The system independently performs time-frequency conversion processing on each monitoring point: the temperature time series of the previous N periods is windowed using a Hanning window function to reduce spectral leakage; the time-domain signal is converted to a frequency-domain representation using a fast Fourier transform algorithm to obtain complex spectral data; the spectral amplitude component is extracted and normalized to generate the amplitude spectral vector for that monitoring point. The amplitude spectral vectors of all monitoring points are combined to form a three-dimensional spectral feature tensor, whose dimensions correspond to spatial location, frequency point, and amplitude value, respectively.

[0064] The dominant disturbance band identification algorithm processes the aforementioned spectral feature tensor. The system presets a dynamically adjusted amplitude threshold, calculated based on the statistical characteristics of historical spectrum data, typically taking the percentile value of the overall amplitude distribution. The frequency axis is divided into equal-width sub-bands, and the average amplitude value of all monitoring points within each sub-band is calculated. All sub-bands are traversed, comparing the relative magnitude of the average amplitude with the threshold, and marking the indices of all bands exceeding the threshold. The system performs peak clustering processing: connected component analysis is performed on the marked bands, merging bands with consecutive frequency positions into widebands; the energy integral value of each wideband is calculated, i.e., the sum of the areas of amplitude values ​​within the band; the top K widebands with the largest energy integral values ​​are selected as the dominant disturbance bands, with the K value configured according to system resource constraints. The identification result includes the center frequency, bandwidth, and peak amplitude information of the dominant disturbance band, stored in a shared memory area.

[0065] The PID parameter fuzzy rule membership function reconstruction module is activated. The system loads the current version of the fuzzy rule library, which contains fuzzy rule sets for the three output terms: proportional coefficient, integral coefficient, and derivative coefficient. The reconstruction objective is set to suppress the influence of the identified dominant disturbance frequency band, specifically achieved by adjusting the distribution characteristics of the membership function. The system analyzes the location distribution of the dominant disturbance frequency band in the control frequency domain: when the dominant disturbance frequency band is located in the low-frequency region, the weight of the integral term rule in the fuzzy set of the rate of change of deviation is increased; when it is located in the mid-frequency region, the width of the membership function of the proportional term rule is adjusted; when it is located in the high-frequency region, the regularization factor of the derivative term rule is modified. The membership function reconstruction adopts a parameterized adjustment strategy: the triangular membership function achieves morphological changes by modifying the vertex coordinate positions; the Gaussian membership function adjusts the mean and standard deviation parameters. The system establishes a mapping relationship table between frequency domain characteristics and control parameters, queries this table to obtain reconstruction parameters based on the characteristics of the dominant disturbance frequency band, and performs online updates of the membership function geometry. The reconstructed fuzzy rule base is immediately loaded into the real-time control thread, replacing the original rule base version.

[0066] The temperature anomaly detection process operates independently and in parallel. The system calculates the temperature change acceleration at each monitoring point in each control cycle. Acceleration data is derived from the time series of temperature values: the temperature values ​​at the current moment and the two previous sampling moments are recorded, and the instantaneous acceleration value is calculated using the second-order central difference formula. The acceleration data from all monitoring points form a spatial distribution matrix, with matrix elements corresponding to the acceleration value at each location in the current period. The system maintains an acceleration time series database, storing the acceleration matrices for the most recent M periods in a circular queue structure, where the value of M is set according to the anomaly evolution timescale.

[0067] The time series clustering analysis module starts its processing cycle periodically. The system extracts the latest data cube from the acceleration database, whose dimensions include spatial coordinates, time points, and acceleration values. A distributed clustering algorithm is used to process the multidimensional data: each monitoring point is treated as an independent time series object, containing acceleration observations from its most recent M periods; a dynamic time warping algorithm is selected to calculate the similarity distance between sequences; density clustering is applied to group the monitoring points, with the cluster radius parameter adaptively adjusted according to the acceleration level. The clustering results output the category label and cluster number of each monitoring point. The system calculates the cluster center value for each cluster: the time dimension mean of the acceleration sequences of all monitoring points within the cluster is calculated to form a representative acceleration curve; the absolute integral area of ​​this curve is calculated as the center value quantification index. All clusters are traversed, and the magnitude of the center value quantification index is compared; the cluster with the largest center value is selected as a candidate anomaly region. The spatial continuity verification step filters discrete noise points: morphological dilation is performed on the candidate clusters to fill the gaps in the monitoring point distribution; the area of ​​connected components is calculated, and fragmented regions with areas smaller than a threshold are removed. The final identified continuous spatial region is marked as the temperature anomaly region, and its boundary coordinate set and center position coordinates are output. Anomaly region information is updated to the system status register in real time for use by the compensation mechanism. The entire detection process, while ensuring real-time performance, effectively distinguishes between normal fluctuations and abnormal states through time series feature analysis, providing accurate region location for subsequent control correction.

[0068] Example 3: See Figure 4 The execution process of the multi-level compensation mechanism triggered by the temperature anomaly region begins with a spatial coordinate mapping operation. The system reads the set of boundary coordinates of the temperature anomaly region, which is represented in the device coordinate system. The compensation node network is predefined on the surface of the semiconductor heater, and each node has a unique physical location identifier and logical address. The positioning algorithm calculates the Euclidean distance between the center point of the anomaly region and all compensation nodes: Let the coordinates of the center point of the anomaly region be... , No. The coordinates of each compensation node are: Then the distance for:

[0069]

[0070] System screening meets The set of nodes ( (For the preset radius of action), select the smallest one from them. The corresponding node serves as the nearest compensation node, and outputs its logical address to the control bus.

[0071] The feedforward compensation channel activation process is executed based on the node address. Each compensation node is associated with an independent feedforward control channel, which includes a digital signal processor, a digital-to-analog converter module, and a power amplifier. The activation command is transmitted to the target node via the control bus, triggering the following operations: initializing the signal processor's memory space; configuring the sampling rate of the digital-to-analog converter module (matching the dominant disturbance frequency band); and preheating the power amplifier to operating temperature. The channel readiness status is fed back to the central controller via a status register.

[0072] The correction signal synthesis module operates based on the characteristics of the dominant disturbance frequency band. The system reads the parameters of the identified dominant disturbance frequency band: center frequency. ,bandwidth and reference range Signal generation employs a phase inversion strategy: correction signal

[0073]

[0074] in: Amplitude adjustment factor (based on) (Scaling) The phase angle of the disturbance signal is tracked in real time (extracted from the temperature signal via Hilbert transform). This indicates a phase reversal operation. Signal synthesis is completed in the digital domain, and after digital-to-analog conversion, it is injected into the feedforward compensation channel. The dynamic signal amplitude adjustment mechanism monitors the average temperature deviation in abnormal temperature areas in real time and scales it proportionally. The value is matched to the disturbance intensity.

[0075] The construction of the multi-dimensional evaluation index system was carried out concurrently. Steady-state error index. The calculation method is as follows: after the temperature enters a steady-state phase (with fluctuations less than ±0.5℃ for 10 consecutive cycles), the average absolute value of the temperature deviation over the last 20 sampling cycles is taken. Overshoot index. The acquisition method is to record the maximum peak value during the temperature rise process. With the set target value percentage of deviation Adjusting time indicators Defined as: from the moment control is initiated to the moment the temperature first reaches its maximum. The data is collected over a period of five cycles. All metrics are recorded via a dedicated data acquisition channel with timestamp accuracy down to the millisecond level.

[0076] Weighted fusion generates a comprehensive control effectiveness score The formula is:

[0077]

[0078] in: Preset weighting coefficients (satisfying) ), These are the maximum allowable thresholds for each indicator. (Score value) Normalized to the [0,1] interval, the higher the value, the better the control performance.

[0079] The historical case database uses a structured data format for storage. Each record contains: a timestamp field (accurate to microseconds) and a control parameter combination field (storage ratio coefficient). Integral Time Differential time The current value), performance rating field (records) (Value). A spatiotemporal composite index is created for the data table: a B+ tree index is used for the time dimension to achieve fast range queries; a multi-dimensional R-tree index is used for the parameter dimension to support similarity retrieval. Stored procedures ensure data integrity through transaction logs, and redundant checks are performed on each write.

[0080] The phase correction mechanism and performance evaluation work in tandem through event synchronization. When the compensation channel is activated, the system initiates a correction effect monitoring window: for the next 30 control cycles, it collects average temperature data for the temperature anomaly area every second. If the standard deviation of the average temperature from the set value decreases by more than 15%, the correction is marked as effective, and the compensation parameter database is automatically updated; if the expected result is not achieved, a compensation gain adjustment is triggered: the correction signal amplitude is increased in steps of 0.1. The process continues until the standard deviation decrease threshold is met or the amplitude upper limit is reached. The entire implementation process ensures timing determinism through a hardware interrupt mechanism, and the compensation signal injection is strictly synchronized with the PID control cycle to avoid signal conflicts. The evaluation data stream employs a dual-buffer design: a front-end buffer acquires raw data in real time, while a background processing thread performs index calculations and storage operations, ensuring that the real-time performance of the control system is unaffected by the evaluation process.

[0081] Example 4: See Figure 5 A process formula change event triggers a historical case database retrieval mechanism. The system monitors the formula identifier field of the production management interface. When an update to this field is detected, the feature vector of the new formula is immediately extracted. The feature vector contains elements such as the target temperature curve morphology encoding, heating rate level, and spatial uniformity requirements, and is converted into a 128-dimensional feature descriptor. The case database retrieval engine initiates a similarity matching process, using an improved distance calculation algorithm to search for similar cases in historical records. The algorithm first performs feature dimensionality reduction processing, compressing the 128-dimensional vector to 16-dimensional core features through principal component analysis; then, it calculates the Euclidean distance in the multidimensional vector space of historical cases; finally, it returns the historical record with the smallest distance value as the matching case. A timeout protection mechanism is set in the matching process, limiting the retrieval to 200 milliseconds.

[0082] The control parameter combination from the matching case is extracted and injected into the adaptive PID controller. The parameter combination includes numerical sets of the proportional coefficient KP, integral coefficient KI, and derivative coefficient KD. The system loads the parameter set into the controller's real-time processing unit via register write operations, overwriting the current operating parameters. This parameter set serves as the preheating initialization baseline, and the controller immediately switches to the new parameters. The preheating phase lasts for a fixed duration, typically set to five percent of the complete process cycle, allowing the system to enter a transition state.

[0083] The deviation between real-time temperature data and the set temperature trajectory drives the gradient fine-tuning process. In each control cycle, the system collects the actual temperature values ​​of all monitoring points and calculates the deviation vector between these values ​​and the corresponding set temperature. This deviation vector is then processed using a spatially weighted average to generate a global deviation scalar value. The fine-tuning algorithm iteratively optimizes based on this scalar value: when the absolute value of the deviation exceeds a set threshold, a parameter correction process is initiated; the correction direction is determined by the sign of the deviation—positive deviations correspond to parameter reduction adjustments, and negative deviations correspond to incremental adjustments; the correction magnitude employs an adaptive step-size strategy, with an initial step size of 0.5% of the parameter value, which is then dynamically reduced based on the rate of change of the deviation. The proportional coefficient KP, integral coefficient KI, and derivative coefficient KD are each fine-tuned independently, and the adjusted parameter set takes effect immediately.

[0084] Energy consumption distribution data acquisition is achieved through a distributed power metering system. High-precision power sensors are installed in each power supply branch of the semiconductor heater to collect instantaneous power values ​​at a frequency of ten times per second. The power data from all branches are combined to form a spatial energy consumption distribution heat map, with the heat map grid aligned with the temperature monitoring point array. Heat dissipation status data acquisition includes two independent sources: fan speed data is obtained by reading pulse frequency signals from the fan drive module and converting them into speed values; coolant temperature data is acquired through a thermocouple array embedded in the cooling pipes, with collection points distributed at the inlet, outlet, and key heat exchange areas.

[0085] The Pareto front optimization model for temperature control accuracy and energy consumption is constructed as follows. The system defines two objective functions: the first objective is the temperature control accuracy evaluation function, calculating the root mean square error of all monitoring points; the second objective is the total energy consumption evaluation function, calculating the Joule dissipation per unit time. Constraints include the algebraic relationships after discretization of the thermodynamic balance equations, and the maximum cooling capacity boundary of the heat dissipation system. A multi-objective genetic algorithm is used to solve the model, with a population size of fifty individuals, each representing a set of PID parameter combinations. The evolutionary iteration is performed for one hundred generations, and the solution set is filtered through non-dominated sorting and crowding calculation. Finally, the Pareto optimal solution set is output, and the minimum and maximum values ​​of the proportional coefficient KP, integral coefficient KI, and differential coefficient KD for all parameter combinations in the solution set are extracted to form the PID parameter constraint boundary. See Table 1.

[0086] Table 1: Storage structure data table of historical case library.

[0087]

[0088] The parameter constraint boundary application mechanism is activated with each PID parameter update. When the fine-tuned proportional coefficient KP exceeds the boundary range, the system automatically clamps it to the nearest boundary value; the integral coefficient KI and derivative coefficient KD undergo the same processing. Boundary values ​​are updated every 30 seconds, and Pareto optimization model calculations are performed through a background thread to ensure that constraints evolve dynamically with the system state. After the process formulation switches to the steady-state stage, the fine-tuning step size decays exponentially until the adjustment stops, allowing the parameter set to converge to a stable value. The newly stable parameter set is combined with the current formulation feature code to generate a new case, which is then added to the historical case database after verification.

[0089] Example 5: Real-time monitoring of the change in the thermal expansion coefficient of a semiconductor heater is achieved using a distributed sensor network. A micro-strain sensor array is embedded at key structural locations within the heater. Each sensor indirectly acquires the material's thermal expansion characteristics by measuring local deformation. The system periodically reads sensor data, with the sampling frequency synchronized with the temperature control cycle. The thermal expansion coefficient change calculation module receives the raw deformation data and first performs temperature drift compensation: referencing the temperature value at the corresponding location in the current temperature field distribution matrix to eliminate the influence of pure thermal expansion effects; then, it calculates the effective mechanical deformation value and converts it into an instantaneous thermal expansion coefficient value using a material constitutive model. This instantaneous value is compared with the initial calibration reference value, and the absolute change is calculated and stored in a circular buffer. The system records the time series of changes at each monitoring point, with the series length covering the entire process cycle.

[0090] When the change in the coefficient of thermal expansion exceeds the tolerance threshold, the system activates the thermal deformation compensation algorithm. The tolerance threshold is dynamically configured based on the heater material type; a strict threshold is set for brittle materials, while a lenient threshold is set for tough materials. The threshold determination employs a triple verification mechanism: first, it checks whether the change at a single monitoring point exceeds the threshold; second, it verifies whether adjacent monitoring points simultaneously exceed the limit; and finally, it confirms whether the change trend continues for three sampling cycles. When all conditions are met, the system sends a hardware interrupt signal to the compensation control unit.

[0091] The thermal deformation compensation algorithm is implemented using finite element numerical simulation. The algorithm preloads a 3D structural mesh model of the heater, with mesh nodes spatially aligned with the temperature monitoring points. Input parameters include the current temperature field distribution matrix, a library of material elastic modulus parameters, and structural constraint boundary conditions. The calculation process solves the thermo-mechanical coupling equations: calculating the thermal stress distribution based on the temperature gradient; deriving the deformation displacement field through the stress-strain relationship; and inversely calculating the temperature field distortion based on the deformation displacement. The compensation amount is output as a 2D matrix, with matrix row numbers corresponding to monitoring point numbers, and matrix element values ​​representing the required temperature compensation value at that location. The compensation calculation is completed on a dedicated coprocessor, ensuring results are output within two hundred milliseconds.

[0092] The compensation superposition operation is implemented at the output of the adaptive PID controller. The system is configured with a multi-channel signal mixer, receiving two input signals: the main channel is the original power regulation signal from the PID controller, and the auxiliary channel is the compensation matrix output by the thermal deformation compensation algorithm. Before superposition, signal standardization processing is performed: the compensation matrix is ​​converted to power regulation dimensions and scaled according to the heater power response coefficient; the main channel signal is buffered and delayed for alignment to eliminate the impact of algorithm calculation delay. Superposition adopts a point-by-point addition principle: each power control channel independently receives the compensation amount at its corresponding monitoring point and algebraically adds it to the original output signal. The final mixed control signal is transmitted to the power actuator.

[0093] The material parameter adaptive update mechanism operates continuously. The system establishes a material performance degradation model and records the change in the coefficient of thermal expansion each time compensation is activated. When the change in a single instance reaches seventy times the upper limit of the threshold, the material is determined to have entered a fatigue state, and the elastic modulus parameter in the material database is automatically updated. The parameter update uses a moving average algorithm: the new elastic modulus value is equal to the weighted fusion result of the historical average and the current measurement value, with the three most recent measurements having a higher weight. The updated parameters are immediately applied to the next compensation calculation.

[0094] The interruption recovery process ensures system robustness. When a process interruption occurs during compensation activation, the system freezes the compensation amount: saving the current compensation matrix to non-volatile memory; disconnecting the compensation signal output channel; and maintaining the original output of the PID controller. After process recovery, the system checks the pre-interruption state: if the interruption time is less than five minutes, the frozen compensation amount is directly restored and continues to accumulate; if it exceeds five minutes, the thermal expansion coefficient monitoring process is re-executed, and the compensation amount is generated based on the latest data. Throughout the compensation cycle, the temperature monitoring system continuously tracks the compensation effect. When it detects that the target area temperature has returned to the normal range for more than ten control cycles, the compensation channel is automatically shut down and the monitoring flag is reset.

[0095] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0096] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for temperature control of a semiconductor heater based on adaptive PID, characterized in that, include: Real-time acquisition of temperature data from multiple monitoring points of a semiconductor heater to construct a temperature field distribution matrix; A set temperature trajectory is generated based on the temperature field distribution matrix, and the heating power of the semiconductor heater is adjusted in a closed loop using an adaptive PID controller. During the closed-loop control process, the fuzzy rule base of PID parameters is dynamically reconstructed based on the temperature fluctuation trend. When an abnormal temperature region is detected, a multi-level compensation mechanism is triggered to perform phase correction on the adaptive PID controller; The fuzzy rule base for PID parameters is dynamically reconstructed based on temperature fluctuation trends, including: Extract the spectral characteristics of temperature fluctuations over N consecutive sampling periods; The frequency bands whose amplitude exceeds a threshold in the spectral characteristics of the temperature fluctuations are identified as the dominant perturbation frequency bands. The fuzzy rule membership function of PID parameters is reconstructed with the goal of suppressing the dominant disturbance frequency band; Triggering a multi-level compensation mechanism to perform phase correction on the adaptive PID controller, including: Locate the nearest compensation node based on the spatial coordinates of the temperature anomaly area; Activate the feedforward compensation channel of the nearest compensation node; A correction signal with the opposite phase to the dominant disturbance frequency band is injected into the feedforward compensation channel.

2. The semiconductor heater temperature control method based on adaptive PID according to claim 1, characterized in that, Generating a set temperature trajectory based on the temperature field distribution matrix includes: Analyze the axial and radial temperature gradients in the temperature field distribution matrix; Retrieve the corresponding temperature distribution template based on the current process formula; By integrating the axial temperature gradient, radial temperature gradient, and temperature distribution template, a set temperature trajectory with time dimension constraints is generated.

3. The semiconductor heater temperature control method based on adaptive PID according to claim 2, characterized in that, The heating power of a semiconductor heater is controlled in a closed loop using an adaptive PID controller, including: Calculate the deviation between real-time temperature data and the set temperature trajectory, as well as the rate of change of the deviation. Based on the deviation and the rate of change of deviation, the fuzzy rule base of PID parameters is queried, and the proportional coefficient correction, integral coefficient correction, and derivative coefficient correction are output. Adjust the real-time PID control parameters according to the proportional coefficient correction, integral coefficient correction, and derivative coefficient correction.

4. The semiconductor heater temperature control method based on adaptive PID according to claim 1, characterized in that, When an abnormal temperature region is detected, it includes: Calculate the acceleration of temperature change at each monitoring point; Time series cluster analysis was performed on the temperature change acceleration; The region with the largest cluster center value is marked as a temperature anomaly region.

5. The semiconductor heater temperature control method based on adaptive PID according to claim 1, characterized in that, Also includes: A multi-dimensional evaluation index system for the temperature control process is established, which includes steady-state error index, overshoot index, and settling time index. The multidimensional evaluation index system is weighted and fused to generate a comprehensive control effectiveness score; The comprehensive control performance score is combined with the current control parameters and stored in the historical case library.

6. The semiconductor heater temperature control method based on adaptive PID according to claim 5, characterized in that, Also includes: When a change in process formulation is detected, the most similar historical control parameter combination is retrieved from the historical case database; The historical control parameter combination is injected into the adaptive PID controller for preheating initialization; The preheating parameters are finely adjusted based on the deviation between the real-time temperature data and the set temperature trajectory.

7. The semiconductor heater temperature control method based on adaptive PID according to claim 6, characterized in that, Also includes: Collect energy consumption distribution data and heat dissipation status data of semiconductor heaters, construct a Pareto front optimization model for temperature control accuracy and energy consumption, and output PID parameters to constrain the boundary through the Pareto front optimization model.

8. The semiconductor heater temperature control method based on adaptive PID according to claim 7, characterized in that, Also includes: Real-time monitoring of changes in the coefficient of thermal expansion of semiconductor heater materials; When the change in the coefficient of thermal expansion exceeds the tolerance threshold, the thermal deformation compensation algorithm is activated, and the compensation amount output by the thermal deformation compensation algorithm is superimposed on the output of the adaptive PID controller.

Images