A high-precision temperature control method and system for a photolithography machine

Abstract

The application discloses a high-precision temperature control method and system of a photoetching machine, belongs to the technical field of temperature control of photoetching machines, collects a temperature array of a lower surface of a mask, a heat flow density distribution of a bearing surface of an exposure table and a thermal pressure difference signal of an air floating layer, carries out feature coding, and constructs a high-dimensional thermal coupling situation feature matrix; then, a dynamic thermal weight correlation graph is generated based on the matrix, exposure key nodes are taken as anchor points, a graph neural network is used to explicitly model a nonlinear thermal path between nodes; further, a spatial gradient change rate of unit area heat flux of a mask area is calculated according to the thermal weight graph, and is mapped to a temperature control array of the exposure table to generate a nonlinear control function surface; and the power output and response delay parameters of a local heating unit are dynamically adjusted according to the control function, so that the method can realize dynamic thermal field equalization with nanometer-level precision, effectively suppress thermal drift, and improve photoetching imaging quality and yield of advanced processes.

Description

Technical Field

[0001] This invention relates to the field of temperature control technology for lithography machines, and specifically to a high-precision temperature control method and system for lithography machines. Background Technology

[0002] As semiconductor manufacturing advances to 7nm and more advanced process nodes, the imaging distortion problem caused by temperature micro-perturbations in the lithography process is becoming increasingly prominent. Traditional lithography machine temperature control systems mostly rely on single-sensor thermal feedback mechanisms or fixed-time segmented PID control strategies, which are difficult to respond to sub-micron-level thermal drift caused by coupling disturbances such as ambient thermal conductivity drift, local thermal coupling between the mask and wafer, and nonlinear evolution of platform mechanical thermal expansion and contraction. Existing technologies struggle to dynamically model and correct thermally induced offset paths, leading to frequent problems such as focal length micro-shifts, blurred pattern edges, and wafer layer misalignment, severely restricting yield control in extreme limiting processes.

[0003] Especially in double patterning, immersion lithography, and EUV lithography scenarios, the thermal coupling relationships between the lower surface of the mask, the photomask fixing ring, the exposure stage support surface, and its air-bearing system are complex, and microscale thermal disturbances exhibit characteristics such as being unmeasurable, uncontrollable, and asymmetric. Although some systems have attempted to introduce CFD simulation to assist modeling, the difficulty in accurately inputting on-site environmental parameters still prevents the active monitoring and dynamic compensation of multi-source thermal disturbances during operation, greatly limiting the potential for improving lithography accuracy in advanced process lines. Summary of the Invention

[0004] The purpose of this invention is to provide a high-precision temperature control method and system for a lithography machine to address the shortcomings in the prior art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a high-precision temperature control method for a lithography machine, comprising:

[0006] Acquire the time-series thermal response signals of the heat source region inside the lithography machine, including the temperature array on the lower surface of the mask, the micro heat flux density distribution on the exposure stage bearing surface, and the multi-point thermal pressure difference signals of the supporting air-float layer;

[0007] The time-series thermal response signal is feature-encoded to construct a high-dimensional thermal coupling situation feature matrix;

[0008] Based on the thermal coupling feature matrix, a dynamic thermal weight association map is constructed, and the key exposure nodes are used as map anchors. A multi-layer graph neural network is introduced to explicitly model the potential nonlinear thermal paths between nodes.

[0009] Based on the thermal weight map, the spatial gradient change rate of heat flux per unit area on the lower surface of the mask is calculated and mapped to the real-time temperature control unit of the exposure stage to generate the corresponding nonlinear control function surface.

[0010] Based on the nonlinear control function surface, the power output and response delay parameters of the local heating unit of the exposure stage are dynamically adjusted;

[0011] The corrected temperature control response characteristics are fed back into the thermal disturbance characteristic matrix as the initial conditions for the next cycle of thermal modeling, so as to achieve continuous multi-cycle adaptive temperature control.

[0012] Preferably, the feature encoding of the time-series thermal response signal includes:

[0013] The mask temperature array, heat flux density sequence, and thermal pressure difference signal are normalized respectively.

[0014] The variation features of each type of thermal signal within the time window are extracted based on a convolutional variational autoencoder network, with an encoding dimension of no less than 64.

[0015] Multi-scale convolutional kernels are set in the encoder to capture short-term high-frequency disturbances and long-term low-frequency thermal drift patterns respectively.

[0016] The latent variable distribution of each type of heat source signal is generated through a reparameter sampling mechanism to reflect its potential thermal evolution trend.

[0017] Preferably, the construction of the high-dimensional thermal coupling situation feature matrix includes:

[0018] The latent variables are concatenated according to the data source to form a three-dimensional input tensor, with the dimensions in the order of signal type, time step and feature encoding length.

[0019] The input tensors are weighted and fused to generate a weighted graph of the interaction relationships between heat sources;

[0020] A node interconnection weight matrix is ​​constructed based on the graph, where nodes represent spatial locations and edges represent thermal impact paths.

[0021] The final result is a thermal coupling situation characteristic matrix, where the element values ​​represent the propagation intensity and coupling trend of thermal disturbances between different nodes.

[0022] Preferably, constructing a dynamic thermal weight correlation map based on the thermal coupling feature matrix includes:

[0023] Each row or column in the high-dimensional thermal coupling situation feature matrix is ​​regarded as a heat source node;

[0024] The initial weights of edges are determined using a formula for calculating the feature similarity between nodes, where the similarity calculation is based on cosine similarity or Euclidean distance.

[0025] Construct an undirected weighted graph structure for all nodes and edges, where the edge weights represent the propagation intensity of thermal disturbances;

[0026] The edge weights are periodically updated using a time-sliding window mechanism to form a dynamic heat map.

[0027] Preferably, the exposure key nodes are used as graph anchor points, and a multi-layer graph neural network is introduced to explicitly model the potential nonlinear thermal paths between nodes, including:

[0028] Through lithography path planning and wafer layout analysis, multiple heat-sensitive nodes, including the central region of the exposure mask, the geometric center of the wafer, and the outermost circular region, were identified.

[0029] Heat-sensitive nodes are set as anchor nodes in the heat weight graph, and anchor nodes serve as the core of information aggregation and propagation in the graph;

[0030] Construct a graph convolutional neural network with no fewer than two layers, each layer including adjacency matrix normalization, feature aggregation, and non-linear activation operations;

[0031] The input is a dynamic thermal weight correlation map and the thermal feature vectors of each node, and the output is a high-order thermal response embedding representation of each node.

[0032] Use ReLU or LeakyReLU as the activation function to maintain nonlinear modeling capability;

[0033] Residual connections and Dropout mechanisms are used to prevent deep network degradation and overfitting.

[0034] Preferably, calculating the spatial gradient rate of change of heat flux per unit area on the lower surface of the mask includes:

[0035] Based on the embedding features of the nodes corresponding to the mask region in the thermal coupling feature matrix, a spatial distribution model of heat flux is established by combining the node spatial coordinates.

[0036] A two-dimensional Gaussian fitting method was applied to the model to obtain the continuous field of heat flux;

[0037] The first derivatives of the heat flux per unit area in the X and Y directions are calculated between adjacent grid cells using a numerical difference algorithm.

[0038] The derivative results are used to construct a heat flux gradient field, which is used to represent the directionality and rate of change of thermal perturbation on the mask surface.

[0039] Preferably, generating the corresponding nonlinear control function surface includes:

[0040] The input thermal gradient values ​​and their historical response data of each temperature control unit are input into the spline interpolation module; a nonlinear mapping function surface between the thermal input and the temperature control output is constructed using a three-dimensional B-spline interpolation algorithm.

[0041] Preferably, dynamically adjusting the power output and response delay parameters of the local heating unit of the exposure stage includes:

[0042] The adjustment of the power output includes: determining the target power of each local heating unit based on the Z-axis power value of each control point in the nonlinear control function surface; and introducing a time factor to perform a weighted average of the historical power change rate.

[0043] The adjustment of the response delay parameter includes: establishing a first-order inertial thermal response model for each local heating unit; dynamically adjusting the control delay time of each unit according to the current heat flux gradient magnitude and direction change rate; and applying an advance amount to the heating command through a time prediction compensation algorithm to offset the thermal response lag.

[0044] The present invention also provides a high-precision temperature control system for a lithography machine, comprising:

[0045] Signal acquisition module: Acquires the time-series thermal response signals of the heat source region inside the lithography machine, including the temperature array on the lower surface of the mask, the micro heat flux density distribution on the exposure stage bearing surface, and the multi-point thermal pressure difference signals of the supporting air-float layer;

[0046] Feature encoding module: Encodes the time-series thermal response signal to construct a high-dimensional thermal coupling situation feature matrix;

[0047] Association Graph Construction Module: Based on the thermal coupling feature matrix, a dynamic thermal weight association graph is constructed, and key exposure nodes are used as graph anchor points. A multi-layer graph neural network is introduced to explicitly model the potential nonlinear thermal paths between nodes.

[0048] Calculation module: Based on the thermal weight map, calculate the spatial gradient change rate of heat flux per unit area on the lower surface of the mask, and map it to the real-time temperature control unit of the exposure stage to generate the corresponding nonlinear control function surface;

[0049] Adjustment module: Based on the nonlinear control function surface, dynamically adjust the power output and response delay parameters of the local heating unit of the exposure stage;

[0050] Feedback module: Feeds back the corrected temperature control response characteristics to the thermal disturbance feature matrix as the initial condition for the next cycle of thermal modeling, so as to achieve continuous multi-cycle adaptive temperature control.

[0051] The technical effects and advantages provided by the present invention in the above technical solution are as follows:

[0052] 1. This invention achieves adaptive modeling and dynamic control of the complex thermal field inside a lithography machine by introducing a multi-cycle feedback mechanism based on the thermal disturbance feature matrix. Compared to traditional single-cycle open-loop or fixed-parameter control methods, this invention can correct the modeling results using temperature control response characteristics after each control cycle, ensuring that the thermal modeling process remains consistent with the actual operating state of the equipment. This effectively suppresses the cumulative errors caused by environmental fluctuations, thermal inertia, and load differences, thereby achieving higher-precision temperature control.

[0053] 2. This invention further combines nonlinear control function surfaces with dynamic feedback mechanisms, enabling the local heating unit of the exposure stage to maintain adaptive optimization of power output and response delay parameters during multi-cycle operation. This method not only improves the system's response speed to transient thermal disturbances but also reduces the risk of overcompensation caused by regional thermal coupling, ensuring stable control of focal length, size, and alignment accuracy in lithography machines at 7nm and below processes. Attached Figure Description

[0054] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0055] Figure 1 This is a flowchart of the method of the present invention.

[0056] Figure 2 This is a flowchart of the system modules of the present invention. Detailed Implementation

[0057] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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, 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.

[0058] Example 1, please refer to Figure 1 As shown in this embodiment, a high-precision temperature control method for a lithography machine includes:

[0059] Acquire the time-series thermal response signals of the heat source region inside the lithography machine, including the temperature array on the lower surface of the mask, the micro heat flux density distribution on the exposure stage bearing surface, and the multi-point thermal pressure difference signals of the supporting air-float layer;

[0060] The time-series thermal response signal is feature-encoded to construct a high-dimensional thermal coupling situation feature matrix;

[0061] Based on the thermal coupling feature matrix, a dynamic thermal weight association map is constructed, and the key exposure nodes are used as map anchors. A multi-layer graph neural network is introduced to explicitly model the potential nonlinear thermal paths between nodes.

[0062] Based on the thermal weight map, the spatial gradient change rate of heat flux per unit area on the lower surface of the mask is calculated and mapped to the real-time temperature control unit of the exposure stage to generate the corresponding nonlinear control function surface.

[0063] Based on the nonlinear control function surface, the power output and response delay parameters of the local heating unit of the exposure stage are dynamically adjusted;

[0064] The corrected temperature control response characteristics are fed back into the thermal disturbance characteristic matrix as the initial conditions for the next cycle of thermal modeling, so as to achieve continuous multi-cycle adaptive temperature control.

[0065] To address the characteristics of thermal disturbances in photomasks—characterized by small spatial scales and rapid temporal evolution—this invention integrates an array of MEMS thermocouple sensors within the photomask frame structure. This sensor array is arranged in a rectangular grid pattern, with node spacing controlled within 1.5 mm to ensure coverage of the entire photomask working area. Each sensing node employs a miniature thermocouple chip with a thermal response time of less than 5 milliseconds.

[0066] Sensing data is acquired via a high-speed analog-to-digital converter (ADC) module, with a sampling frequency set above 1000 Hz to ensure the capture of sub-second thermal abrupt events. The acquired temperature data undergoes initial denoising in a local storage buffer before being input into a boundary fitting interpolation algorithm module. This algorithm constructs a family of interpolation functions based on a combination of cubic B-spline interpolation and a two-dimensional Gaussian weighted function, reconstructing a continuous temperature distribution map of the mask surface, thus enabling spatial prediction and filling of the thermal state at non-sensing points.

[0067] To address the heat flux density variations on the wafer support surface, this invention deploys several thin-film heat flux sensors at equal intervals below the quartz panel of the exposure stage. Each sensor group includes a heat flux meter and a thermometer unit stacked vertically. The collected heat flux time-series signal is input to a Fourier inversion decoupling algorithm module. Through frequency domain filtering and principal component analysis, the high-frequency noise components and the low-frequency system thermal inertial response in the signal are separated to obtain the effective heat flux variation function.

[0068] To enhance the spatial correlation of heat flux data, this invention introduces a "wafer movement trajectory matrix" into heat flux modeling. Specifically, based on the wafer displacement trajectory recorded in the exposure stepper log, the heat flux signal at each moment is remapped to spatial matrix coordinates, achieving two-dimensional modeling of heat flux density over time and location. The output is a "micro-heat flux density spatiotemporal distribution model," used to invert the thermal equilibrium state and temperature field variation trend of the wafer region.

[0069] To monitor vertical thermal disturbances during the floating process of the photolithography platform, this invention deploys more than six sets of micro piezoelectric thin-film thermistors along the axis of symmetry of the platform in the air-float support system. Each assembly includes a piezoelectric film, a thermistor bridge, and a differential pressure converter. This assembly can respond in microseconds to pressure fluctuations caused by localized thermal expansion and contraction within the air-float cavity.

[0070] The system acquires the differential pressure signal in real time and uses a "thermal compensation bridge algorithm" to correct for environmental drift and compensate for zero drift caused by changes in background temperature. The differential pressure signal is then fitted into a "thermal differential pressure variation trend function" by a weighted dynamic fitting module to calculate the probability of the platform's vertical floating state and thermal instability.

[0071] Since the data sampling periods, formats, and noise characteristics of the three types of heat source regions are different, in order to ensure the temporal consistency and spatial coupling of the subsequent thermal coupling map construction, this invention introduces a "temporal normalization and sliding window caching mechanism" for unified preprocessing:

[0072] Time series normalization: By setting a unified reference time axis (in milliseconds), linear interpolation or nearest neighbor interpolation is used to align all heat source data to a unified time node.

[0073] Sliding window mechanism: Set a fixed window with a length of 100 milliseconds. Data is sorted and cached in the window by time. The window overlap rate is 50% to smoothly capture high-frequency disturbances.

[0074] Outlier removal: The Z-score method is used to remove noisy data points that are greater than 3 times the standard deviation, ensuring data quality.

[0075] The module ultimately outputs a "uniformly aligned thermal response data frame" as input for subsequent thermal coupling modeling.

[0076] To unify the characteristic representation of various types of thermal signals, this invention first normalizes the raw data of each type. Using time series as the unit, the thermal signal within each sampling window is standardized to a zero-mean, unit-variance format. Let the original thermal signal sequence be X, the standard deviation be σ, and the mean be μ. The normalization result is X minus μ divided by σ, thereby eliminating the problems of dimensional differences and amplitude inconsistencies between different signals.

[0077] The standardized thermal signals will be fed into corresponding Convolutional Variational Autoencoder (CVAE) networks for feature encoding. CVAE is a deep neural network structure based on unsupervised learning, capable of automatically extracting latent variation patterns from the original input and generating compressed representations.

[0078] The overall structure of CVAE includes the following modules:

[0079] Input layer: Receives sequence data of each type of thermal response signal within a fixed-length time window (e.g., 100 sampling points).

[0080] Convolutional encoder: It consists of multiple one-dimensional convolutional layers, using convolutional kernels of different sizes (such as 3, 5, 7) to extract short-term disturbances and long-term trends of thermal signals in parallel. After convolution, batch normalization and activation functions are applied.

[0081] Latent variable sampler: Based on the feature distribution of the encoder output, the mean vector μ and the standard deviation vector σ are extracted. The latent variable z is generated by sampling through the reparameterization technique, such that z = μ + σ × ε, where ε is a standard normal distribution random variable, and z is the compressed feature representation of the thermal signal.

[0082] Decoder: Taking the latent variable z as input, it attempts to reconstruct the original sequence, thereby optimizing the encoder quality through the loss function.

[0083] The CVAE training process employs a joint loss function, including a reconstruction loss term and a Kullback-Leibler (KL) divergence term, which respectively measure the reconstruction accuracy and the degree of standard normal approximation of the latent variable distribution. After training, only the encoder module is deployed into the lithography control system to achieve real-time online encoding of the thermal signal.

[0084] Each type of heat source signal, after being CVAE encoded, will obtain a 64-dimensional latent vector representation, also known as the thermal evolution latent variable vector, which is used to reflect the disturbance trend, frequency pattern and thermal inertial response characteristics of the heat source within the current time window.

[0085] To uniformly manage the propagation behavior of thermal disturbances from multiple heat sources across multiple spatial nodes, this invention further organizes and fuses the aforementioned latent variable vectors. The specific steps are as follows:

[0086] Latent variables extracted from different heat sources (e.g., photomasks, exposure stages, air flotation systems) within the same time window are concatenated to form a three-dimensional tensor, with the following dimension order:

[0087] Dimension 1: Types of thermal signals (3 types in total); Dimension 2: Time step (number of frames in the sliding window); Dimension 3: Encoding dimension (64-dimensional latent variables);

[0088] This invention introduces a weighted fusion method based on an attention mechanism to extract information from the above tensor in both spatiotemporal dimensions.

[0089] The temporal attention layer is used to identify which frame of data within the window contains key perturbations;

[0090] Spatial attention layers are used to identify the intensity of interactions between different heat sources.

[0091] Finally, a set of interaction weight maps between heat sources is obtained, where the weight values ​​represent the synchronicity of thermal response and the coupling strength.

[0092] Each heat source measurement point is considered as a node in the graph, and the edge weights between nodes are the attention weights of the corresponding heat source pairs in the above interaction graph. An adjacency matrix is ​​constructed to record the node connection relationships, and the edge weights are defined as thermal perturbation similarity or heat conduction probability.

[0093] Based on the above heat source node map, a thermal coupling situation feature matrix is ​​further generated. This matrix is ​​an n×n dimensional symmetric matrix, where n is the total number of heat source nodes, and the element value in the i-th row and j-th column represents the thermal disturbance propagation intensity or directional coupling trend between the i-th node and the j-th node in the current time window.

[0094] This thermal coupling situation feature matrix serves as the core input for subsequent graph structure modeling (as shown in the convolutional network), which helps to deeply model the nonlinear coupling mechanism between thermal field changes and spatial conduction paths.

[0095] To characterize the coupling and propagation relationship of thermal disturbances among multiple heat source nodes in a lithography machine, this invention constructs a thermal weight correlation map based on a thermal coupling situation feature matrix. The specific implementation is as follows:

[0096] Node definition: Each row or column in the situation feature matrix is ​​represented as a "heat source node". Each node represents a specific spatial location or device part (such as a point on the lower surface of a mask, a certain area of ​​an exposure stage, or a measuring point of an air flotation system).

[0097] Edge weight calculation: For any two nodes in a thermally coupled relationship, calculate their feature similarity as the edge weight. Similarity can be calculated in one of the following two ways:

[0098] Cosine similarity is calculated by dividing the dot product of the feature vectors of two nodes by their norms.

[0099] The reciprocal form of the Euclidean distance is used to emphasize the differences between features.

[0100] The edge weights are normalized to a range of 0 to 1.

[0101] Graph structure generation: All nodes and their edge weights form an undirected weighted graph structure. The i-th row and j-th column of the adjacency matrix A represents the thermal coupling strength between the i-th node and the j-th node.

[0102] Dynamic update mechanism: Through the sliding time window mechanism, the system recalculates the edge weights after a fixed time period (e.g., 500 milliseconds) to reflect the time-varying nature of thermal perturbations and obtain a dynamic thermal weight spectrum sequence that evolves over time.

[0103] Considering that thermal disturbances do not propagate uniformly in space and that temperature fluctuations in the exposure area have the greatest impact on image accuracy, this invention introduces an anchor point mechanism in the thermal map to enhance the model's ability to model key areas.

[0104] Anchor point identification method: Based on wafer layout and lithography path planning data, the following three types of thermally sensitive areas are extracted as anchor points:

[0105] The center point directly below the mask;

[0106] Wafer geometric center point;

[0107] The outermost ring region at the edge of the wafer.

[0108] Anchor point enhancement strategy: When constructing the heatmap, the edge weights of anchor points and their neighboring nodes are assigned a connectivity enhancement coefficient (e.g., multiplied by 1.5), which enables anchor points to have greater information aggregation capabilities during the propagation of the graph neural network.

[0109] Information dissemination priority: During the training phase, the system prioritizes sampling subgraphs containing anchor points to enhance the model's accuracy in modeling regions highly sensitive to thermal disturbances.

[0110] The anchor point mechanism can significantly improve the ability to capture key thermal paths and avoid local distortion caused by average modeling.

[0111] To address the nonlinear, asymmetric, and time-varying characteristics of thermal disturbance propagation in photolithography systems, this invention further introduces a multi-layer graph neural network model to explicitly model the heat transfer path in the graph structure. The implementation steps are as follows:

[0112] Graph Convolutional Network Structure Construction: A Graph Convolutional Network (GCN) is selected as the basic structure, containing at least two graph convolutional layers. Each layer includes the following substructures:

[0113] The adjacency matrix normalization operation performs symmetric normalization on the node connection relationships to ensure stable propagation;

[0114] Feature aggregation operation: Each node aggregates information from its neighboring nodes at the current layer;

[0115] Non-linear activation functions, such as ReLU or LeakyReLU, enhance the network's fitting ability;

[0116] Residual connections and Dropout mechanisms improve training stability and reduce overfitting.

[0117] Input definition:

[0118] Node input features: thermal evolution latent variable vector (64-dimensional) from the output of the CVAE encoder;

[0119] Adjacency matrix: generated from dynamic heat weight map, reflecting the heat propagation intensity between nodes;

[0120] Output representation: The embedding vector of each node after multi-layer graph convolution, representing its thermal response state in the global thermal field.

[0121] Training Mechanism: The graph neural network is trained using a self-supervised learning mechanism. The training objective is to minimize the mean square error between the actual thermal disturbance path trajectory and the model's predicted path. Training samples are derived from thermal drift trajectories collected from historical temperature control systems, and labeled paths are generated offline.

[0122] Modeling output: The model finally outputs the high-order thermal feature embedding of each node, which can be used as input to the subsequent thermal compensation path generation module, or directly used in scenarios such as focal length correction, thermal balance control, and equipment pre-scheduling.

[0123] First, based on the thermal coupling feature matrix and thermal weight map constructed in the preceding steps, this invention models and analyzes the thermal perturbation state of the mask region. By extracting the graph neural network embedding feature vectors (usually real vectors of length 64) of the corresponding nodes in this region, and combining them with the positional distribution of each node in the spatial coordinate system, a spatial distribution model of heat flux is constructed.

[0124] The specific steps are as follows:

[0125] A two-dimensional Gaussian kernel function is used to weighted fit the thermal intensity of all mask region nodes to obtain a spatially continuous and differentiable heat flux distribution function. This function is defined on the XY plane, using a unit area (e.g., 1 square millimeter) as the computational unit to describe the heat flux intensity at each location.

[0126] The heat flux distribution function described above is subjected to first-order numerical difference operations along the X-axis and Y-axis in the spatial domain. That is, the difference of heat flux values ​​at adjacent grid points is calculated and divided by the relative distance to obtain the gradient vector field. The magnitude of the gradient represents the rate of change of heat flux per unit area, and the direction represents the diffusion trend of thermal disturbance.

[0127] For the mask edge region, a combination of mirror boundary extension and finite difference is used to process the boundary gradient value to avoid local abrupt distortion caused by boundary truncation.

[0128] The final output heat flux gradient field contains the spatial gradient vector information of all nodes under the mask, which serves as the basis for the next step of generating temperature control commands.

[0129] Since the photomask and exposure stage in a photolithography system belong to different spatial coordinate systems, this invention establishes a unified optical spatial mapping model to achieve accurate conversion of the photomask thermal gradient field to the exposure stage temperature control unit.

[0130] The spatial mapping mechanism includes the following steps:

[0131] Using optical system calibration data provided by the lithography machine manufacturer, a two-dimensional affine transformation matrix is ​​constructed between the mask surface and the heating array on the exposure stage. This matrix includes operations such as scaling, rotation, and translation. The transformed thermal gradient field can correspond one-to-one with the physical array position on the exposure stage.

[0132] An affine transformation is applied to each coordinate point in the thermal gradient field to obtain its corresponding position in the exposure stage coordinate system, and the thermal gradient modulus value corresponding to that position is assigned to the corresponding temperature control unit.

[0133] To prevent non-physical high-energy inputs, all thermal gradient values ​​are linearly normalized to the range of 0 to 1, and an upper limit threshold (such as 0.8) is set as the maximum value of the heating power factor to prevent local hot spots from causing thermal excitation disturbances.

[0134] Through the above process, each temperature control unit of the exposure stage can accurately sense the thermal disturbance state of the mask in space, thereby achieving distributed power control.

[0135] To further improve the response sensitivity and stability of the temperature control system to changes in thermal gradient, this invention constructs a nonlinear control function surface based on the mapped gradient input value to guide the power response strategy of each temperature control unit.

[0136] The implementation process is as follows:

[0137] Each exposure stage temperature control unit acts as a control point, receiving the following inputs:

[0138] Current gradient value (after normalization);

[0139] Historical response records (such as power input and local thermal response over the last 10 cycles);

[0140] Using a 3D B-spline interpolation algorithm, the input and output of each control point are mapped to spatial function points, and the whole is fitted into a continuous control function surface defined in the XY plane and the Z-axis represents the power response value.

[0141] The B-spline order can be set to 3 (cubic) to obtain a smooth response;

[0142] The grid resolution can be set to 10×10 or finer to improve local fitting accuracy.

[0143] The curvature of a surface can describe the nonlinear control characteristics of a system. The larger the curvature, the more sensitive the temperature control response in that region, requiring smaller disturbances to achieve faster heating or cooling. The smaller the curvature, the stronger the thermal inertia, requiring smoother temperature control actions.

[0144] The control surface is dynamically updated every 50 milliseconds based on the thermal gradient input. The update process runs in parallel on the control processor (such as FPGA or DSP) to ensure timeliness.

[0145] To prevent the control surface from triggering new feedback oscillations under abnormal thermal disturbances or signal abrupt changes, this invention also introduces the following safety and robustness mechanisms:

[0146] Before the control command is officially issued, the system predicts the thermal field state at the next moment caused by the control function surface based on the fast thermal response model. If the predicted value exceeds the set safe temperature rise threshold of the equipment (such as ±1.2℃), the control gain will be automatically reduced.

[0147] The power variation of the same unit within each cycle is limited to no more than 20% of that in the previous cycle to avoid excessive thermal response.

[0148] The control output is processed by a low-pass filter with a cutoff frequency of 10 Hz to suppress response fluctuations caused by high-frequency noise.

[0149] In this invention, the target power of each heating unit is calculated using a control function surface. The control function surface is defined as a three-dimensional continuous function, with its independent variable being the spatial coordinates (X, Y) of the heating unit, and its function value being the desired output power (unit: watts). This function is constructed using the aforementioned B-spline interpolation method and updated in real time according to the heat flux gradient.

[0150] The specific power output update method includes the following steps:

[0151] Each local heating unit obtains the corresponding function value on the control surface as the current target power based on its coordinate position on the exposure stage, denoted as P. t (x, y).

[0152] To avoid system instability caused by sudden power fluctuations, this invention introduces a power-weighted averaging mechanism. That is:

[0153] Current cycle output power = α × P t (x, y) + (1-α) × P t-1(x, y); where α is a smoothing factor, ranging from 0.2 to 0.5, dynamically adjusted according to response speed requirements. The power calculation cycle is set to 50 milliseconds. Each cycle, a new power command is input to the heating drive unit via a digital-to-analog converter module to update the high-frequency control command. This method can suppress high-frequency oscillations while maintaining control response speed.

[0154] Because the thermal response in real-world systems often exhibits deviations and lags, this invention introduces a closed-loop temperature feedback correction mechanism during heating power adjustment to ensure accurate temperature control. Specifically, this includes:

[0155] At least one high-response thermocouple sensor with a sampling frequency of 1000 Hz is placed near each heating unit to monitor the rate of temperature change in the area.

[0156] Based on the thermal gradient field input, a desired temperature rise trajectory curve is constructed, which is usually a linearly rising curve or an S-shaped gradual rise curve, with a fixed target temperature difference (e.g., 1.5℃) and rise time (e.g., 500 milliseconds).

[0157] The actual rate of temperature change is compared with the expected curve to calculate the error ΔT, and a correction coefficient k = f(ΔT) is constructed from this error. For example, k can be defined as 1 - ΔT / ΔT max ΔT max This is the maximum allowable deviation value for the system, typically set to 0.3℃.

[0158] Multiply the correction factor k by the current target power to obtain the new adjustment power, forming a closed-loop regulation to ensure that the actual temperature curve closely follows the expected change trajectory.

[0159] Considering the thermal inertia lag between the heating unit receiving the power command and the actual temperature change, this invention further introduces a response delay modeling and prediction compensation method.

[0160] The following thermal model is established for each heating unit:

[0161] Temperature change rate = (input power × thermal gain - current temperature) ÷ time constant τ; where the thermal gain is in degrees Celsius / watt, and the time constant τ is the response delay of the unit (in milliseconds), which is obtained by fitting historical heating data, and typically ranges from 150 milliseconds to 500 milliseconds.

[0162] When the control command is executed, the predicted temperature response is estimated one cycle in advance. Based on the current input power and historical conditions, the expected temperature rise is calculated, and power adjustment is applied in advance to offset the delay effect.

[0163] The heating units in different regions are divided into three categories—fast, medium, and slow—based on their response time constant τ. Different advance compensation amounts are used for each category (e.g., 0 milliseconds advance for fast, 50 milliseconds advance for medium, and 100 milliseconds advance for slow) to achieve hierarchical compensation control.

[0164] In photolithography platforms, local heating units often exhibit thermal coupling, and independent control between units may lead to local overcompensation or regional thermal interference. To avoid this problem, this invention proposes an adaptive thermal coupling cooperative control strategy, comprising the following steps:

[0165] Based on the heat flux map and the heat conduction path between nodes, the thermal coupling degree W(i,j) between any two adjacent heating units is calculated. The value ranges from 0 to 1, representing the proportion of conduction influence per unit power in the neighborhood.

[0166] During each power update, the power output P(i) of a certain unit is corrected according to the output of the adjacent units: P(i)_new=P(i)-Σ[W(i,j)×ΔP(j)]; where ΔP(j) is the power change of the adjacent j-th unit relative to the previous cycle.

[0167] Set a maximum total power threshold for the area (e.g., 200 watts). Exceeding this threshold will trigger an overall equalization power reduction operation to prevent local system overload.

[0168] This coordinated adjustment strategy can effectively mitigate the amplification effect of local thermal offset and improve the overall thermal stability of the system.

[0169] In the high-precision temperature control method for a lithography machine provided by this invention, in order to achieve rapid adaptation and stable compensation of the system to minute thermal disturbances, a feedback mechanism for temperature control response characteristics is further proposed. That is, the heating control results and local thermal response of each cycle are fed back to the thermal disturbance feature matrix in a structured form as the starting state for thermal modeling and path prediction in the next control cycle.

[0170] This approach enables continuous multi-cycle adaptive modeling and control closed-loop at the system level, effectively suppressing the accumulation of control deviations caused by thermal inertia, load fluctuations, or environmental disturbances, and improving the long-term stability and robustness of thermal field regulation.

[0171] Temperature control response characteristics refer to the mapping result between the actual output power of each local heating unit and the thermal feedback within the current cycle, mainly including:

[0172] Control execution data: The final power output value of each heating unit, in watts, denoted as P(i), where i is the heating unit number.

[0173] Thermal feedback data: the rate of temperature change in the corresponding temperature zone, in degrees Celsius per second, denoted as dT(i) / dt.

[0174] Dynamic response coefficient: defined as the ratio of temperature rise rate to power, i.e., G(i)=[dT(i) / dt]÷P(i), used to measure local thermal response sensitivity.

[0175] Delay offset: Records the time delay between the actual activation of the system response and the power delivery, in milliseconds, denoted as τ(i), and is calculated using timestamps.

[0176] All the above features form a periodic response feature vector R(i), with the following format: R(i) = [P(i), dT(i) / dt, G(i), τ(i)]; where each component is a real number and is normalized to ensure the consistency of subsequent algorithm inputs.

[0177] To integrate temperature control response features into the existing thermal modeling system, this invention employs the following feature feedback fusion mechanism:

[0178] The original thermal perturbation feature matrix M(t) is a three-dimensional tensor with dimensions of [number of nodes × feature length × time window], and each node contains its thermal signal encoded features.

[0179] After each control cycle ends, the matrix slides one step (e.g., 50 milliseconds) along the time dimension, the earliest frame of data is removed, and the new frame of response data is inserted at the end.

[0180] The response feature vector R(i) generated in the current period is mapped to the thermal signal encoding vector Z(i) corresponding to node i, and updated using the following fusion function: Z(i)_new=β×Z(i)_prev+(1-β)×R(i); where β is the fusion balance factor, and it is recommended to take a value of 0.6 to 0.8.

[0181] This invention forms a continuously updating modeling closed loop through the above feedback method, and its complete control cycle includes the following five stages:

[0182] Thermal data acquisition: Collect multi-source thermal disturbance data of the mask surface, exposure stage and air flotation system in the current cycle.

[0183] Feature encoding and graph modeling: Using CVAE and graph neural networks, the collected data is transformed into node feature vectors and heat propagation paths.

[0184] Control function generation and heating execution: Generate a nonlinear control function surface based on the thermal gradient and output heating power.

[0185] Thermal response feedback extraction: Collect the thermal feedback after heating and calculate the response feature vector.

[0186] Matrix update and modeling starting point correction: The response features are incorporated into the feature matrix as the starting state for the next cycle of modeling.

[0187] By performing state correction on the feature matrix after each control cycle, this system possesses the following capabilities:

[0188] When the response capability of the heating unit decreases due to environmental changes, the system can automatically identify the response offset and adjust the power forecast for the next cycle.

[0189] Long-term thermal offset no longer accumulates, but instead returns to a stable trajectory through feedback correction in each cycle;

[0190] Feedback information, as input features, improves the prediction accuracy of neural networks and avoids the spread of pure prediction errors.

[0191] Example 2, please refer to Figure 2 As shown in this embodiment, a high-precision temperature control system for a lithography machine includes:

[0192] Signal acquisition module: Acquires the time-series thermal response signals of the heat source region inside the lithography machine, including the temperature array on the lower surface of the mask, the micro heat flux density distribution on the exposure stage bearing surface, and the multi-point thermal pressure difference signals of the supporting air-float layer;

[0193] Feature encoding module: Encodes the time-series thermal response signal to construct a high-dimensional thermal coupling situation feature matrix;

[0194] Association Graph Construction Module: Based on the thermal coupling feature matrix, a dynamic thermal weight association graph is constructed, and key exposure nodes are used as graph anchor points. A multi-layer graph neural network is introduced to explicitly model the potential nonlinear thermal paths between nodes.

[0195] Calculation module: Based on the thermal weight map, calculate the spatial gradient change rate of heat flux per unit area on the lower surface of the mask, and map it to the real-time temperature control unit of the exposure stage to generate the corresponding nonlinear control function surface;

[0196] Adjustment module: Based on the nonlinear control function surface, dynamically adjust the power output and response delay parameters of the local heating unit of the exposure stage;

[0197] Feedback module: Feeds back the corrected temperature control response characteristics to the thermal disturbance feature matrix as the initial condition for the next cycle of thermal modeling, so as to achieve continuous multi-cycle adaptive temperature control.

[0198] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A method for high-precision temperature control of a lithography machine, characterized in that: The method comprises the steps of: acquiring time-series thermal response signals of internal heat source regions of a photolithography machine, including a temperature array of a lower surface of a mask, a micro-thermal flow density distribution of a bearing surface of an exposure table, and a multi-point thermal pressure difference signal of a support air float layer; encoding the time-series thermal response signals to construct a high-dimensional thermal coupling situation feature matrix; constructing a dynamic thermal weight correlation graph based on the high-dimensional thermal coupling situation feature matrix, taking key exposure nodes as anchor points of the graph, and introducing a multi-layer graph neural network to explicitly model potential nonlinear thermal paths between nodes; calculating a spatial gradient change rate of unit area heat flux of the lower surface of the mask based on the dynamic thermal weight correlation graph, and mapping it to real-time temperature control units of the exposure table to generate corresponding nonlinear control function surfaces; dynamically adjusting power output and response delay parameters of local heating units of the exposure table according to the nonlinear control function surfaces; feeding back corrected temperature control response features to the thermal disturbance feature matrix as initial conditions for thermal modeling in the next period to realize adaptive temperature regulation in continuous multiple periods.

2. The method for high-precision temperature control of a photoetching machine according to claim 1, wherein: Wherein, the feature encoding of the time-series thermal response signals comprises: normalizing the temperature array of the mask, the thermal flow density sequence, and the thermal pressure difference signal respectively; extracting the change features of each type of thermal signal within a time window based on a convolutional variational auto-encoding network, and the encoding dimension is not less than 64; setting multi-scale convolution kernels in the encoder to capture short-term high-frequency disturbances and long-term low-frequency thermal drift patterns respectively; generating hidden variable distributions of each type of heat source signal through a reparameterization sampling mechanism to reflect the potential thermal evolution trend.

3. The high-precision temperature control method for a lithography machine according to claim 2, characterized in that: Wherein, the construction of the high-dimensional thermal coupling situation feature matrix comprises: splicing the hidden variables according to data sources to form a three-dimensional input tensor, and the dimension order is signal type, time step, and feature encoding length; performing weighted fusion on the input tensor to generate a weight graph of the interaction relationship between heat sources; constructing a node interconnection weight matrix based on the graph, taking nodes to represent spatial positions and edges to represent thermal influence paths; finally forming a thermal coupling situation feature matrix, and the matrix element value represents the propagation strength and coupling trend of thermal disturbances between different nodes.

4. The high-precision temperature control method for a lithography machine according to claim 3, characterized in that: The construction of the dynamic thermal weight correlation graph based on the high-dimensional thermal coupling situation feature matrix comprises: regarding each row or column in the high-dimensional thermal coupling situation feature matrix as a heat source node; determining the initial weight of the edge by using a node feature similarity calculation formula, and the similarity calculation is based on cosine similarity or Euclidean distance; constructing a non-directional weighted graph structure for all nodes and edges, wherein the edge weight represents the propagation strength of thermal disturbances; periodically updating the edge weight through a time sliding window mechanism to form a dynamic thermal weight correlation graph.

5. The high-precision temperature control method for a lithography machine according to claim 4, characterized in that: Wherein, taking the key exposure nodes as anchor points of the graph, and introducing a multi-layer graph neural network to explicitly model potential nonlinear thermal paths between nodes comprises: determining multiple heat-sensitive nodes including a central region of an exposure mask, a geometric center of a wafer, and an outermost circular region through lithography path planning and wafer layout analysis; setting the heat-sensitive nodes as anchor nodes in the dynamic thermal weight correlation graph, and the anchor nodes are the core of information aggregation and propagation in the graph. Constructing a graph convolutional neural network containing no less than two layers, each layer containing an adjacency matrix normalization, a feature aggregation, and a nonlinear activation operation; The input is a dynamic heat weight correlation graph and a heat feature vector of each node, and the output is a high-order heat response embedding representation of each node; Using ReLU or LeakyReLU as an activation function to maintain nonlinear modeling capability; Preventing deep network degradation and overfitting through residual connection and Dropout mechanism.

6. The method of claim 5, wherein the temperature of the lithography machine is controlled with a high precision. Calculating the spatial gradient change rate of the unit area heat flux on the lower surface of the mask, including: Based on the embedding features of the nodes corresponding to the mask area in the high-dimensional heat coupling situation feature matrix, a heat flux spatial distribution model is established combined with the spatial coordinates of the nodes; Applying a two-dimensional Gaussian fitting method to the model to obtain a continuous heat flux field; Using a numerical difference algorithm to calculate the first-order derivative of the unit area heat flux in the X and Y directions between adjacent grid cells; The derivative results constitute a heat flux gradient field, which is used to represent the directionality and change rate of thermal disturbance on the mask surface.

7. The high-precision temperature control method for a lithography machine according to claim 6, characterized in that: Wherein, Generating a corresponding nonlinear control function surface, including: Inputting the input heat gradient values of each temperature control unit and its historical response data into the spline interpolation module; using a three-dimensional B-spline interpolation algorithm to construct a nonlinear mapping function surface between heat input and temperature control output.

8. The method of claim 7, wherein the temperature of the lithography machine is controlled with a high precision. Dynamically adjusting the power output and response delay parameters of the local heating units of the exposure table, including: The adjustment of the power output includes: determining the target power of each local heating unit according to the Z-axis power value of each control point in the nonlinear control function surface; introducing a time factor to weight and average the historical power change rate; The adjustment of the response delay parameter includes: establishing a first-order inertial heat response model for each local heating unit; dynamically adjusting the control delay time of each unit according to the current heat flux gradient amplitude and direction change rate; applying an advance amount to the heating instruction through a time prediction compensation algorithm to offset the heat response lag.

9. A high-precision temperature control system of a lithography machine, for implementing the high-precision temperature control method of any one of claims 1-8, characterized in that: Including: Signal acquisition module: acquiring time-series thermal response signals of the internal heat source area of the lithography machine, including the temperature array of the lower surface of the mask, the micro-heat flux density distribution of the exposure table bearing surface, and the multi-point thermal pressure difference signal of the support air floating layer; Feature encoding module: feature encoding of the time-series thermal response signal to construct a high-dimensional heat coupling situation feature matrix; Correlation graph construction module: based on the high-dimensional heat coupling situation feature matrix, a dynamic heat weight correlation graph is constructed, the key nodes of exposure are taken as the anchor points of the graph, and a multi-layer graph neural network is introduced to explicitly model the potential nonlinear thermal path between nodes; Calculation module: according to the dynamic heat weight correlation graph, the spatial gradient change rate of the unit area heat flux on the lower surface of the mask is calculated and mapped to the real-time temperature control unit of the exposure table to generate a corresponding nonlinear control function surface; Adjustment module: dynamically adjusting the power output and response delay parameters of the local heating units of the exposure table according to the nonlinear control function surface; Feedback module: feeding back the corrected temperature control response features to the thermal disturbance feature matrix as the initial condition for the next period of thermal modeling to realize continuous multi-period adaptive temperature regulation.

Images