An evolutionary reasoning method for multi-stage computational lithography driven by a physical world model

Abstract

This invention discloses a physical world model-driven multi-stage computational lithography evolution reasoning method, which can unfold the lithography state in the potential state space of each stage and, combined with process intervention strategies, achieve high-precision modeling and optimization of the lithography pattern evolution process. The method includes the following steps: Step 1, constructing a complete dataset by collecting lithography data under different layout and process parameter combinations; Step 2, mapping the physical state of each stage of the lithography process to the potential space, and modeling the potential state of each stage using a spatial embedding method; Step 3, establishing a process intervention strategy model in the potential space, and planning process adjustment schemes for different stages; Step 4, establishing a state transition model to simulate the continuous evolution of the potential state between stages, predicting the potential state of the next stage through the current stage potential state and intervention strategy, and achieving continuous deduction of the lithography state by iteratively updating the intervention scheme.

Description

Technical Field

[0001] This invention belongs to the field of semiconductor manufacturing technology, specifically relating to an evolutionary reasoning method for multi-stage computational lithography driven by a physical world model. Background Technology

[0002] Photolithography is a core process in modern integrated circuit manufacturing, and its precision and reliability directly affect chip yield, performance, and manufacturing costs. With the increasing complexity and integration of chip designs, advanced process nodes (such as 7nm, 5nm, and below) place higher demands on photolithography. Key steps in the photolithography process include chip layout design, mask generation, optical imaging, photoresist exposure, and image formation after development. Each step is governed by different physical mechanisms, such as optical diffraction, optical proximity effect, material chemical reactions, and process fluctuations. To improve photolithography precision and reduce manufacturing defect rates, computational lithography is widely used in current technologies. Computational lithography typically simulates the photolithography process using physical or empirical models, thereby predicting and optimizing the imaging effect during the mask design stage. Its core methods mainly include: (1) Optical Proximity Correction (OPC): Optical proximity correction is a traditional technique for addressing pattern distortion problems in photolithography. Due to the non-ideal characteristics of optical diffraction and projection systems, small features on the mask often exhibit linewidth deviations or edge curvature after exposure. The OPC method makes the actual exposed pattern closer to the desired pattern by locally fine-tuning the mask features. Existing OPC techniques are mainly divided into regularized OPC and model-driven OPC. Regularized OPC corrects mask features through predefined geometric rules, such as adding auxiliary lines and correcting sharp corners. This method has low computational cost and is easy to implement, but its effectiveness is limited in complex layouts and high-density designs. Model-driven OPC is based on optical and chemical physics models and iteratively optimizes the mask geometry to make the developed pattern meet the design specifications. This method has high accuracy, but high computational cost, and the computational cost increases exponentially with the increase in design scale. Although OPC can improve the accuracy of local patterns, its adaptability under multi-stage photolithography processes, complex layouts, and continuous process interventions is still limited.

[0003] (2) Inverse Lithography Technology (ILT): Inverse lithography is a method that directly generates an ideal mask through numerical optimization. ILT uses the development result of the target pattern as the optimization objective and uses an iterative algorithm to reverse-engineer the mask shape that yields the desired pattern. The advantage of ILT technology lies in its ability to provide more accurate mask designs for complex patterns, sharp corners, and high-density regions, making it more advantageous than the traditional OPC method at advanced process nodes. However, the ILT method usually requires high-precision physical simulation as the basis for optimization evaluation, resulting in high computational costs and numerous iterations, making it difficult to apply efficiently in large-scale chip design. In addition, the ILT method still has limitations in multi-stage lithography evolution, continuous process intervention, and multi-objective optimization, making it difficult to directly handle the dynamic changes throughout the lithography process.

[0004] (3) Data-driven and machine learning methods: With the development of big data technology, some technical solutions attempt to collect a large amount of layout, mask and exposure result data and use machine learning models for lithography prediction and optimization. Such methods can speed up the prediction process and reduce the reliance on traditional physical simulators. Data-driven methods usually include supervised learning and deep learning models, which can learn the mapping relationship between the input layout and the development result, thereby quickly predicting the photoresist image or ADI (After Development Image) effect.

[0005] Although machine learning methods have achieved some success in static prediction, they have the following shortcomings: the models can usually only handle single-stage input and output and cannot capture the multi-stage continuous evolution of the lithography process; the implicit and continuous characteristics of process intervention are difficult to incorporate into model training, resulting in unreliable predictions under complex process adjustments; and the model's generalization ability is limited, making it difficult to adapt to the needs of large-scale and diversified chip design.

[0006] In summary, existing technologies have limitations in both improving lithography accuracy and computational efficiency: traditional OPC and ILT methods involve large computational loads, making it difficult to efficiently handle large-scale chip designs; while data-driven methods offer high computational efficiency, they struggle to simulate continuous process evolution and multi-stage state changes. Therefore, this invention, based on existing lithography theory and computational lithography methods, and combined with the physical characteristics and process requirements of multi-stage lithography processes, proposes a novel modeling and optimization method to overcome these technical shortcomings and provide efficient and reliable lithography process support for chip manufacturing at advanced process nodes. Summary of the Invention

[0007] To overcome the shortcomings of existing technologies, the present invention aims to provide a multi-stage computational lithography evolution reasoning method driven by a physical world model, which can unfold the lithography state in the potential state space of each stage and, combined with process intervention strategies, achieve high-precision modeling and optimization of the lithography pattern evolution process.

[0008] To achieve the above objectives, the present invention can be implemented using the following specific technical solutions: The aforementioned evolutionary reasoning method for multi-stage computational lithography driven by a physical world model includes the following steps: Step 1: By collecting lithography data under different layout and process parameter combinations, construct a complete dataset. The dataset contains the photoresist development results of each sample under specific exposure dose, focal length offset, threshold and illumination distribution. Step 2: Map the physical states of each stage of the photolithography process to the latent space, and model the latent states of each stage using the spatial embedding method; Step 3: Establish a process intervention strategy model in the potential space and plan process adjustment schemes for different stages; Step 4: Establish a state transition model to simulate the continuous evolution of the potential state between stages. Predict the potential state of the next stage by using the potential state of the current stage and the intervention strategy, and realize the continuous deduction of the lithography state by iteratively updating the intervention scheme.

[0009] Furthermore, in step 1, when constructing the dataset, the multi-stage lithography state of each sample is unfolded in the order of layout-mask-photoresist image-developed image, retaining the process parameters and local physical information corresponding to each stage; and each sample in the dataset is calibrated at multiple levels, including local feature standardization, field of view unification and physical parameter normalization.

[0010] Furthermore, the specific content of step 2 includes: Step 2.1, Latent Space Mapping and Stage Evolution Modeling: A pre-trained and frozen encoder network is used. The first step in the photolithography process The original physical state of each stage The mapping is represented as a low-dimensional latent feature vector, and the latent state change vectors of adjacent stages are defined, with the change vectors implying the first... Specific physical dynamic information at each stage; Step 2.2: Construction of the physical information latent space based on principal component analysis: Constructing a physical information latent subspace. To constrain The direction of change; that is, introducing a set of orthogonal and normalized basis vectors to construct the basis matrix. The set of process intervention directions is defined as the physical information latent space of the basis matrix. .

[0011] Furthermore, the specific content of step 3 includes: Step 3.1, Stochastic Modeling of Intervention Strategies: Treating process interventions as random variables and using conditional probability distributions to model the process interventions, the resulting intervention coefficients are obtained. ; Step 3.2, Physical constraint mapping of intervention vectors: Mapping the intervention coefficients obtained from direct sampling... Mapped to the constructed physical information latent space In this process, the mapping is performed through the basis matrix. In practice, the vector span of the intervention coefficients forms the intervention direction subspace for the current stage.

[0012] Furthermore, in step 3.1, regarding the first... Each stage involves building a strategy model. The input to the strategy model includes the potential state of the current stage. and process window parameters The strategy model first extracts high-dimensional features from the input using a visual transformer to obtain conditional features of the fusion state and process constraints. The feature is then input into a multilayer perceptron, which outputs the Gaussian distribution parameters—mean—following the potential intervention coefficients. with standard deviation By introducing randomness, the model can generate diverse intervention candidates in the potential space and adaptively adjust the intervention distribution according to the evolution of the current state, supporting multi-step planning and exploration.

[0013] Furthermore, in step 4, for the same process batch, the first... Phase 1 The potential state of the step is The corresponding process intervention is A diffusion converter is used as the transfer model. The backbone network is used to learn the dominant dynamics of state evolution; the diffusion transformer processes the input triples. Joint feature extraction is performed to output a general intermediate representation. After the diffusion transform, a stage-aware multilayer perceptron is introduced. Based on the current stage index, the corresponding MLP branch is dynamically selected to map the general representation to the state space of the specific stage, thereby generating the predicted state for the next time step. .

[0014] Furthermore, in step 4, after obtaining the predicted state... Afterwards, it is necessary to assess whether the current intervention has driven the state to approach the target distribution of this stage; let the first... The true target state at each stage is By freezing the encoder and mapping it to the latent space, The evolutionary loss function is defined as the Euclidean distance between the predicted state and the target state, and the loss function is calculated for the intervention vector. gradient norm g t .

[0015] Furthermore, an experience threshold is set. As a criterion for determining whether the evolution within a stage has reached local stability: like This indicates that the current intervention has not yet fully driven state evolution and predicted the state. The policy model fed back to the same stage triggers a new round of intervention generation and state prediction; like If the current state has entered a locally stable region, then the predicted state is accepted. This serves as the final output of this stage and propels the process to the next stage.

[0016] Compared with the prior art, the present invention has the following advantages: (1) Unlike traditional lithography modeling methods, this invention enables seamless integration between multiple stages, from layout design to mask generation, and then to the formation of photoresist images and post-development images (ADI), ensuring that the physical dynamics of each stage closely align with the requirements of subsequent stages. By learning the latent space of stage-specific physical information and introducing a contrastive variational optimization paradigm, this invention achieves joint learning of process intervention and state transitions without intermediate supervision. This method does not rely on traditional explicit physical modeling or black-box methods that overly depend on sample data. Instead, it optimizes the entire lithography process through inherent physical constraints and evolutionary relationships, thereby ensuring the accuracy and reliability of the results while maintaining high efficiency.

[0017] (2) This invention can handle cross-node lithography tasks, not only performing well under existing process nodes but also adapting to changes in different technology nodes, demonstrating excellent cross-process capability. By introducing an innovative contrastive variational optimization paradigm, this invention breaks through the limitations of traditional methods, enabling dynamic adjustment of lithography process intervention through the potential space of physical information, thereby maximizing the optimization effect. The application of this technology not only promotes the cutting-edge development of the lithography simulation field but also provides strong technical support for the automated design, process optimization, and rapid iteration of the lithography process, and has broad engineering application prospects. Attached Figure Description

[0018] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0020] This invention addresses the multi-stage physical evolution and process intervention optimization problem in photolithography, proposing a physical world model-driven multi-stage computational photolithography evolution reasoning method. This method can unfold the photolithography state in the potential state space of each stage and, combined with process intervention strategies, achieve high-precision modeling and optimization of the photolithography pattern evolution process. For example... Figure 1 As shown, this invention can realize multi-step potential state unfolding within each stage and intervention perception decision-making between each stage. It includes the following steps: (I) Step 1: Construction of the photolithography “layout-mask-photoresist image-developed image (ADI)” dataset. This step constructs a complete dataset by collecting photolithography data under different combinations of layout and process parameters. The dataset contains the photoresist development results for each sample under specific exposure dose, focal length offset, threshold, and illumination distribution. Each sample corresponds to a local region and is standardized to unify the physical field of view, thereby ensuring data consistency and usability.

[0021] This invention systematically collects and organizes lithography data to form a lithography dataset covering different layouts and process parameter combinations, which can be used for multi-stage modeling and process intervention optimization. When constructing the dataset, the multi-stage lithography states of each sample are unfolded in the order of layout-mask-photoresist image-adaptive image (ADI), while retaining the corresponding process parameters and local physical information for each stage. By collecting 280k paired samples under different process conditions, the dataset can comprehensively reflect the characteristics of multi-stage state evolution during the lithography process. The dataset also provides complete local regional physical information, enabling subsequent potential spatial mapping and intervention strategy training to fully utilize the continuous change patterns between states at each stage, providing a reliable data foundation for process intervention.

[0022] Furthermore, this invention performs multi-level calibration on each sample in the dataset, including local feature standardization, field-of-view unification, and physical parameter normalization, thereby ensuring the consistency and generalization ability of the trained model under different process conditions. These processes guarantee that the model can effectively constrain multi-stage evolution when predicting lithography states and planning process interventions, and provide a reliable potential spatial mapping basis for subsequent strategy models.

[0023] (II) Step 2: Latent Space Mapping and Physical Embedding. Based on the dataset from Step 1, this step maps the physical states of each stage of the lithography process to a physical information latent space, and models the latent states of each stage using a spatial embedding method. The latent space not only captures the continuously changing information of each stage, but also constrains the direction of process intervention, ensuring that subsequent interventions conform to physical constraints. By modeling the potential changes of adjacent stages, the physical consistency of the lithography process can be effectively guaranteed, providing a basis for strategy planning.

[0024] This invention encodes the physical states of each stage of the photolithography process into low-dimensional representations through mapping of physical information latent space, enabling multi-stage state modeling and process intervention optimization within the latent space. In this step, the photolithography state of each stage is mapped into a latent vector by an encoder, and stage-specific physical information is embedded in the latent space. Latent space mapping not only captures the continuous physical evolution of each stage of photolithography but also provides constraints on the direction of potential intervention, ensuring that intervention operations are performed within physically permissible limits.

[0025] (1) Latent space mapping and stage evolution modeling: A pre-trained and frozen encoder network is used. The first step in the photolithography process The original physical state of each stage The mapping is represented as a low-dimensional latent feature vector, and its mathematical expression is: , in, Representative stage Down The low-dimensional latent feature vector. To quantify the inter-stage evolution driven by process physics mechanisms, the latent state change vectors of adjacent stages are defined as follows: , in, This represents the sample index, and the change vector implies the stage. Specific physical dynamic information; Representative stage Next Sample The low-dimensional latent feature vector; Representative stage Next Sample The low-dimensional latent feature vector.

[0026] (2) Construction of physical information latent space based on principal component analysis: In order to ensure that subsequent process intervention operations always follow the laws of physical evolution, this invention proposes to construct a physical information latent subspace. To constrain The direction of change. Specifically, a set of orthogonal, normalized basis vectors is introduced to construct the basis matrix. ,in, And satisfy Therefore, the set of permissible process intervention directions is defined as the physical information latent space of this basis matrix. : , in, Is with The relevant scalar coefficients.

[0027] (III) Step 3: Construction of Process Intervention Strategy Model. A process intervention strategy model is established in the latent space to plan process adjustment schemes at different stages. The strategy model generates different intervention directions by sampling the latent states and ensures the feasibility of the intervention schemes through constraints. Through multi-step iteration, the strategy model can explore the evolution path of intervention schemes in the latent space, guide actual process intervention operations, and achieve controllable optimization of multi-stage lithography states.

[0028] In the intelligent control of photolithography processes, this invention further proposes a process intervention planning method based on a stochastic strategy, in order to address the potential physical information space. The body generates diverse intervention programs that conform to physical laws.

[0029] (1) Stochastic Modeling of Intervention Strategies: Given that multiple feasible intervention paths may exist under the same process conditions, this invention treats process intervention as a random variable and models it using conditional probability distribution. Specifically, for the first... Each stage involves building a strategy model. Its input includes the potential state of the current stage. and process window parameters The model first extracts high-dimensional features from the input using a Vision Transformer (ViT) to obtain conditional features representing the fusion state and process constraints. The feature is then input into a multilayer perceptron (MLP), which outputs the Gaussian distribution parameters—mean—following the potential intervention coefficients. with standard deviation ,Right now: .

[0030] To achieve end-to-end differentiable sampling, this invention employs a reparameterization technique. Intervention coefficients Generated by the following formula: , in, This represents a random noise vector. This transformation ensures that the sampling process is relative to the distribution parameters. and Differentiability allows the policy model to be optimized via gradient backpropagation. By introducing stochasticity, the model can generate diverse intervention candidates in the latent space and adaptively adjust the intervention distribution according to the evolution of the current state, supporting multi-step planning and exploration.

[0031] (2) Physical constraint mapping of intervention vectors: intervention coefficients obtained by direct sampling It may not satisfy the first The physical evolution laws at each stage. To ensure that the planned intervention always remains within physically permissible limits, this invention maps it to the aforementioned constructed potential space of physical information. The mapping process is performed through the basis matrix. This is achieved by stretching the column vectors into a subspace of feasible intervention directions for this stage. The final intervention vector after mapping is expressed as: , By organically combining the above-mentioned random modeling and physical constraint mapping, this invention achieves flexible, compliant and optimizable generation of process interventions during the multi-stage evolution of photolithography.

[0032] (iv) Step 4: State Transition and Evolution Model. This step establishes a state transition model to simulate the continuous evolution of the potential state between stages. The potential state of the next stage is predicted using the current stage's potential state and intervention strategy, and the intervention scheme is iteratively updated to achieve continuous deduction of the lithography state. The state transition model supports closed-loop optimization and ensures that the state at each stage satisfies the laws of physical evolution, achieving stable control of the multi-stage lithography process. (1) Stage-aware state prediction: For the same process batch, define the stage-aware state prediction. Phase 1 The potential state of the step is The corresponding process intervention is This invention employs a diffusion transformer (DiT) as the transfer model. The backbone network is used to learn the dominant dynamics of state evolution. The Diffusion Transformer (DiT) first processes the input triplet... Joint feature extraction is performed to output a general intermediate representation. Considering the heterogeneity of feature distribution and numerical scale across different lithography stages, this invention introduces a stage-aware multilayer perceptron (MLP) head after the diffusion transform (DiT). This mechanism dynamically selects the corresponding MLP branch based on the current stage index, mapping the general representation to the state space of a specific stage, thereby generating the predicted state for the next time step. Its formal expression is as follows: .

[0033] (2) Intra-stage evolution stability criterion and iterative control: after obtaining the predicted state Afterwards, it is necessary to assess whether the current intervention has driven the state to approach the target distribution of this stage. Let the first... The true target state at each stage is It is obtained by mapping the frozen encoder to the latent space. The stage evolution loss function is defined as the Euclidean distance between the predicted state and the target state: .

[0034] To quantify the contribution of the current intervention to the state evolution, the loss function is calculated for the intervention vector. Gradient norm: , in, Relative to The gradient. This gradient norm reflects the sensitivity of intervention adjustments to reducing state bias. This invention sets an empirical threshold. As a criterion for determining whether the evolution within a stage has reached local stability: like This indicates that the current intervention has not yet fully driven state evolution and predicted the state. The policy model fed back to the same stage triggers a new round of intervention generation and state prediction; like If so, it is determined that the current state has entered a locally stable region, and acceptance is granted. This serves as the final output of this stage and propels the process to the next stage.

[0035] To prevent infinite iterations within a single stage, this invention sets the maximum number of expansion steps to 10. This mechanism ensures that in multi-step planning, the state transition process follows physical evolution laws and has quantifiable convergence criteria, providing a reliable decision-making basis for closed-loop process control.

[0036] (v) Experimental verification To verify the effectiveness of the proposed multi-stage evolution modeling and intelligent intervention method for lithography processes, experiments were conducted using a system architecture comprising a VAE-based encoding module, a ViT-based policy module, a DiT-based state transition module, and a physical constraint module. A three-stage end-to-end training strategy was employed, with the loss function integrating MSE, BCE, Dice, and edge loss terms. The AdamW optimizer (initial learning rate 1e-6, weight decay 0.01) was used to train for 10 epochs on eight NVIDIA A100 GPUs. The following embodiments are all based on this training configuration.

[0037] (1) Effect verification based on known process conditions This embodiment is used to verify the prediction accuracy and physical consistency of the present invention within a known process parameter space. To this end, a lithography dataset covering the entire process from "layout-mask-photoresist image-developed image" was constructed based on a 55 nm process node. During data generation, the combination and variation of four key process parameters were comprehensively considered: the light source type included three shapes: ring, circle, and bullseye; three different values ​​were selected for the photoresist threshold; the focal length was set to two levels: 0 nm and 50 nm; and the exposure dose was selected to two magnifications: 1.0× and 1.2×. These parameter combinations resulted in 36 differentiated process configurations. The training set contained 280,000 paired samples, covering all 36 process configurations; the domain-specific test set contained 20,000 samples, evenly distributed across all process configurations.

[0038] In the three tasks of mask prediction, photoresist image prediction, and post-development image prediction, the mean intersection-over-union (IoU) ratio obtained using the method of this invention reaches 0.89 to 0.92, the edge F1 score reaches 0.92 to 0.95, and the edge placement error is controlled within the range of 1 to 4 nm. Taking the post-development image prediction task as an example, under typical process configurations, the mean IoU reaches 0.92, the edge F1 score reaches 0.95, and the average edge placement error is 2.3 nm. Experimental results show that this invention, through continuous modeling of the physical evolution of the entire photolithography process, can achieve accurate prediction of multi-stage states.

[0039] To further verify the conformity of the photolithography rules of the image generated by this invention, a commercial photolithography rule checking tool was used to detect defects in the predicted photoresist image. Five process conditions were selected for testing, denoted as A+ to E+. Specifically, A+ condition: bullseye light source, photoresist threshold 0.0923125, focal length 0 nm, exposure dose 1.2×; B+ condition: bullseye light source, photoresist threshold 0.143666, focal length 0 nm, exposure dose 1.2×; C+ condition: ring light source, photoresist threshold 0.0923125, focal length 50 nm, exposure dose 1.0×; D+ condition: circular light source, photoresist threshold 0.1436665, focal length 0 nm, exposure dose 1.0×; E+ condition: circular light source, photoresist threshold 0.0923125, focal length 0 nm, exposure dose 1.0×. The frequency of occurrence of key defect types (Pinch, Bridge, EPE) was statistically analyzed, and the results showed that the prediction results of this invention were highly consistent with the simulation results of commercial tools. Taking the A+ condition as an example, the number of Pinch violations was 12, the number of Bridge violations was 3, and the number of EPE violations was 5, with differences from the simulation results of commercial tools all within 5%. High consistency was also maintained under other test conditions. Furthermore, the prediction deviation of this invention remained stable under different process parameter fluctuations, demonstrating that its modeling capability for fine-grained geometric deformation can meet the accuracy requirements of downstream manufacturing analysis.

[0040] (2) Verification of generalization ability based on unknown process conditions This embodiment is used to verify the generalization ability of the present invention under unknown process parameters and different process nodes. To this end, two out-of-domain test sets were constructed: the first test set is based on the same 55 nm node, using a process combination not involved in training (denoted as F+: ring light source, photoresist threshold 0.119340, focal length 0 nm, exposure dose 1.0×); the second test set uses a publicly available 28 nm process node dataset to evaluate the model's transferability across process nodes. All out-of-domain tests were performed directly on the pre-trained model described in Example 1, without involving additional fine-tuning.

[0041] On the F+ configuration, the edge placement error achieved by this invention is 1 to 2 nm, with an average of 1.5 nm; the mean cross-union ratio reaches 0.88. This result indicates that this invention does not simply memorize the input-output mapping relationship under a specific process configuration, but rather achieves stable extrapolation between different parameter combinations within the same process node by learning the process intervention evolution direction constrained by physical laws.

[0042] On a dataset with a 28 nm process node, this invention maintains high prediction accuracy, with a mean intersection-over-union ratio of 0.89 and edge placement error stable within the 1 to 2 nm range, averaging 1.8 nm. This advantage is attributed to the invention's ability to model the impact of physical interventions at different technology nodes: by learning the intrinsic evolutionary relationship between process interventions and lithography states, the model can accurately capture effective state changes in the current environment when facing unknown process nodes, demonstrating good cross-node generalization ability.

[0043] In summary, this invention achieves accurate modeling and intelligent intervention of the multi-stage evolution of photolithography processes by constructing a potential space constrained by physical information and combining stochastic strategy modeling and state transition prediction. Experimental results show that this invention can obtain high-precision prediction results under both known and unknown process conditions, and the prediction results are highly consistent with simulations using commercial tools. The edge placement error can be controlled within the range of 1 to 4 nm, providing a reliable technical solution for the intelligent control of photolithography processes.

[0044] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A physics world model driven multi-stage computational lithography evolutionary reasoning method, characterized in that, Includes the following steps: Step 1: By collecting lithography data under different layout and process parameter combinations, construct a complete dataset. The dataset contains the photoresist development results of each sample under specific exposure dose, focal length offset, threshold and illumination distribution. Step 2: Map the physical states of each stage of the photolithography process to the latent space, and model the latent states of each stage using the spatial embedding method; Step 3: Establish a process intervention strategy model in the potential space and plan process adjustment schemes for different stages; Step 4: Establish a state transition model to simulate the continuous evolution of the potential state between stages. Predict the potential state of the next stage by using the potential state of the current stage and the intervention strategy, and realize the continuous deduction of the lithography state by iteratively updating the intervention scheme.

2. The method of claim 1, wherein, In step 1, when constructing the dataset, the multi-stage lithography state of each sample is unfolded in the order of layout-mask-photoresist image-developed image, retaining the process parameters and local physical information corresponding to each stage; and each sample in the dataset is calibrated at multiple levels, including local feature standardization, field of view unification and physical parameter normalization.

3. The method of claim 1, wherein, The specific content of step 2 includes: Step 2.1, latent space mapping and phase evolution modeling: adopt a pre-trained and frozen encoder network , the original physical state of the first stage in the lithography process is mapped to a low-dimensional latent feature vector, and the latent state change vector of the adjacent stage is defined, and the change vector contains specific physical dynamic information at the stage ​ Step 2.

2. Constructing the latent subspace of physical information based on principal component analysis , for constraining the change direction of process interventions ; in particular, a set of orthonormal basis vectors is introduced to compose a basis matrix , whose span defines the latent subspace of physical information , thus, denotes the set of all process intervention directions.

4. The method of claim 3, wherein, The specific content of step 3 includes: Step 3.1, Random modeling of intervention strategy: The process intervention is considered as a random variable and modeled using conditional probability distribution, resulting in intervention coefficients ; Step 3.2, Physical constraint mapping of intervention vector: Intervention coefficients directly sampled are mapped to the constructed physical information latent subspace Step 3.2, Physical constraint mapping of intervention vector: Intervention coefficients directly sampled are mapped to the constructed physical information latent subspace Step 3.2, Physical constraint mapping of intervention vector: Intervention coefficients directly sampled are mapped to the constructed physical information latent subspace Step 3.2, Physical constraint mapping of intervention vector: Intervention coefficients directly sampled are mapped to the constructed physical information latent subspace 5. The evolutionary reasoning method for multi-stage computational lithography driven by a physical world model according to claim 4, characterized in that, In step 3.1, for the first Each stage involves building a strategy model. The input to the strategy model includes the potential state of the current stage. and process window parameters The strategy model first extracts high-dimensional features from the input using a visual transformer to obtain conditional features of the fusion state and process constraints. The feature is then input into a multilayer perceptron, and the output latent intervention coefficients follow the mean. with standard deviation By introducing randomness through a Gaussian distribution, the model can generate diverse intervention candidates in the latent space and adaptively adjust the intervention distribution according to the evolution of the current state, supporting multi-step planning and exploration.

6. The evolutionary reasoning method for multi-stage computational lithography driven by a physical world model according to claim 4, characterized in that, In step 4, for the same process batch, the first... Phase 1 The potential state of the step is The corresponding process intervention is A diffusion converter is used as the transfer model. The backbone network is used to learn the dominant dynamics of state evolution; Diffusion transformer on input triplets Joint feature extraction to output a general intermediate representation; introduce a stage-aware multi-layer perceptron head after the diffusion transformer to dynamically select the corresponding MLP branch according to the current stage index, map the general representation to the state space of the specific stage, and thus generate the predicted state at the next time .

7. The evolutionary reasoning method for multi-stage computational lithography driven by a physical world model according to claim 6, characterized in that, In step 4, after obtaining the predicted state... Afterwards, it is necessary to assess whether the current intervention has driven the state to approach the target distribution of this stage; let the first... The true target state at each stage is By freezing the encoder and mapping it to the latent space, The evolutionary loss function is defined as the Euclidean distance between the predicted state and the target state, and the loss function is calculated for the intervention vector. gradient norm g t .

8. The method of claim 7, wherein, Set an empirical threshold θ > 5 x 10 -3 As a criterion for determining whether the evolution within a phase reaches local stability: If g t > 0, it indicates that the current intervention has not sufficiently driven the state evolution, and the predicted state is fed back to the same stage's policy model, triggering a new round of intervention generation and state prediction; If g t ≤ θ, then it is determined that the current state has entered the local stable region, and the predicted state is accepted as the final output of this stage, and the process advances to the next stage.

Images