Physical constraint world model driven glass substrate defect prediction and process optimization method

Abstract

The application discloses a physical constraint world model driven glass substrate defect prediction and process optimization method, comprising: acquiring process data; establishing a thermal-electric-force correlation mechanism model; performing multi-working condition numerical solving based on the thermal-electric-force correlation mechanism model to construct a training sample; constructing and training an action conditional world model, introducing a supervised loss, a physical loss constraint, a timing consistency loss and a confidence evaluation loss; deploying in an online inference system, receiving an action conditional input, predicting a next state and a future multi-step state, outputting a temperature field, a stress field, a displacement or a warping state and a defect risk, and giving a process optimization suggestion. The application improves the prediction ability and physical consistency of the world model for unobserved working conditions through physical constraint embedding and generalization enhancement training, and realizes rapid prediction of glass substrate defect risk and interactive process optimization.

Description

Technical Field

[0001] This invention relates to the technical field of defect prediction and optimization in glass substrate manufacturing processes, and particularly to a method for defect prediction and process optimization of glass substrates driven by a physical constraint world model. Background Technology

[0002] With the rapid development of advanced semiconductor packaging, TGV (Through Glass Via) interconnects, and high-density multi-chip integration technologies, glass substrates are gradually replacing traditional organic substrates and silicon interposers, becoming the core carrier for next-generation advanced packaging, due to their excellent electrical insulation properties, low dielectric loss, high flatness, and thermal expansion coefficient similar to silicon. However, glass substrates involve complex thermo-electrical-mechanical multi-physics coupling processes during via fabrication, copper plating, heat treatment, electrical testing, and subsequent packaging, making them prone to defects such as cracks, warping, delamination, copper layer abnormalities, interface failures, and localized thermal damage.

[0003] The aforementioned defects are often not caused by a single factor, but are the result of the combined effects of thermal stress, chemical stress, and mechanical stress. Among these, thermal stress is one of the most significant defect inducing factors, mainly stemming from the following aspects: mismatch in the coefficient of thermal expansion (CTE) between the glass substrate and the metal layer (such as the copper layer); rapid heating and cooling cycles during the manufacturing process; Joule heating generated when the conductive metal layer is energized; uneven local heat diffusion caused by complex TGV geometry; and residual stress introduced during copper plating or electroplating processes.

[0004] Currently, the control of defects in glass substrate manufacturing processes mainly relies on two methods: post-processing inspection and high-fidelity numerical simulation. For post-processing inspection, technologies such as AOI (Automated Optical Inspection), X-ray / CT (X-ray / Computed Tomography), SEM (Scanning Electron Microscopy), electrical performance testing, and warpage detection can identify defects that have already formed after the process is completed, such as surface cracks, internal voids, delamination, copper layer anomalies, and conductivity failures. However, these inspection methods are post-processing quality control methods and cannot predict whether a set of parameters will lead to high-temperature gradients, high thermal stress, residual stress accumulation, or interface delamination risks during the process parameter setting stage. Therefore, they cannot provide preventative intervention before defects form.

[0005] In high-fidelity multiphysics numerical simulation, existing thermo-electric, thermo-mechanical, or thermo-electric-mechanical coupled simulation methods can analyze temperature, stress, or displacement fields, revealing the physical mechanisms of defect formation to some extent. However, existing technologies have the following shortcomings: First, existing models usually focus on specific physical fields or local coupling relationships, lacking a defect formation mechanism model that unifies and correlates current density, Joule heat source, transient temperature field, CTE mismatch thermal strain, residual effects of copper plating, structural constraints, local stress concentration, and interface failure, making it difficult to establish a continuous physical transmission relationship between process operation conditions and defect risks. Second, for glass substrate systems containing TGV hole arrays, copper layers, interface layers, and clamping boundaries, even if a multiphysics model is established, spatial discretization and time integration are still required using the finite element method, finite volume method, or finite difference method. When performing large-scale operating condition scans for different glass sizes, hole diameters, hole spacing, copper thickness, heating and cooling curves, current densities, copper plating parameters, and clamping methods, the computation time and resource consumption increase significantly, making it difficult to meet the needs of rapid process evaluation and real-time optimization.

[0006] In recent years, data-driven models (such as neural networks) have been widely used in engineering prediction. However, if ordinary data-driven models are trained using only limited simulation samples, the models may fit the training conditions well, but their generalization ability is insufficient for unseen glass dimensions, unseen hole arrays, extreme temperature profiles, new copper plating parameters, or different clamping constraints. More importantly, ordinary neural networks lack physical constraints such as heat propagation, stress balance, and interface failure. When training samples are insufficient or input conditions deviate from the training distribution, prediction results that do not conform to physical laws are prone to occur, such as discontinuous temperature field changes, shifted stress concentration locations, and mismatches between defect risks and stress levels.

[0007] In summary, the existing technologies for glass substrate defect prediction and process optimization have the following technical problems that urgently need to be solved: (1) Defect detection is biased towards post-processing and lacks pre-process prediction capabilities; (2) Multiphysics models lack a unified expression of the thermo-electric-mechanical correlation mechanism for defect prediction; (3) Relying solely on numerical simulation is time-consuming and difficult to support interactive process optimization; (4) Pure data-driven models lack physical constraints and have insufficient generalization capabilities; (5) Traditional process optimization lacks interactive prediction capabilities, and engineers cannot see the impact of different action conditions on future states in real time when adjusting process parameters. Summary of the Invention

[0008] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method for glass substrate defect prediction and process optimization driven by a physical constraint world model, so as to solve the technical problems in the prior art, such as defect detection being too reactive, high time cost of multiphysics simulation, lack of physical constraints and insufficient generalization ability of pure data-driven models, and lack of interactive prediction ability in traditional process optimization.

[0009] To achieve the above objectives, the technical solution provided by this invention is as follows:

[0010] Physically constrained world model-driven methods for glass substrate defect prediction and process optimization include:

[0011] Obtain process data for the glass substrate;

[0012] A thermo-electric-mechanical correlation mechanism model is established based on the process data. The thermo-electric-mechanical correlation mechanism model is used to describe the correlation between the current density field and Joule heat source, transient heat propagation, CTE mismatch thermal strain, residual effect of copper plating, thermoelastic stress response, local stress concentration, and interface crack or delamination failure.

[0013] Using the process data as variable operating condition input, multi-condition numerical solutions are performed based on the thermo-electric-mechanical correlation mechanism model to obtain numerical solution results; and training samples are constructed based on the numerical solution results.

[0014] An action-conditional world model is constructed, and the model is trained offline using the training samples. During the training process, supervised loss, physical loss constraint, temporal consistency loss, and confidence evaluation loss are introduced. The model parameters are updated through joint training of supervised and physical constraints. The physical loss constraint is constructed based on the thermo-electric-mechanical correlation mechanism model.

[0015] The trained action conditional world model predicts the next state and future multiple states of the glass substrate, and provides optimization suggestions for process parameters based on the prediction results.

[0016] Furthermore, the thermo-electric-mechanical correlation mechanism model includes an electric field-Joule heat source sub-model, a transient heat propagation sub-model, a CTE mismatch thermal strain sub-model, a copper plating residual effect sub-model, a thermoelastic constitutive relation sub-model, a mechanical equilibrium-displacement-warping sub-model, and a local stress concentration and interface failure sub-model.

[0017] The electric field-Joule heat source sub-model is constructed as follows:

[0018] First, solve for the current density. :

[0019] ;

[0020] For electric field strength, For electric potential, The conductivity varies with temperature; Represents the electric potential Gradient calculation;

[0021] The governing equation for electric potential is:

[0022] ;

[0023] For spatial differential operators;

[0024] For a Joule heat source within a conductive region, it is represented as:

[0025] ;

[0026] When expressed in terms of resistivity and current density, the Joule heat source is written as:

[0027] ;

[0028] Resistivity;

[0029] The transient heat propagation sub-model is expressed as follows:

[0030] ;

[0031] For density, For specific heat capacity, Thermal conductivity, For temperature, As a Joule heat source, Other internal heat sources besides Joule heating;

[0032] CTE mismatch thermal strain quantum model: for the first For this type of material, the thermal strain is expressed as:

[0033] ;

[0034] in, For the first The coefficient of thermal expansion of similar materials, For reference temperature, It is a unit tensor; when the coefficient of thermal expansion is approximately constant within the temperature range, it simplifies to:

[0035] ;

[0036] First calculate the first The thermal strain of these materials is used to describe the free expansion trends of the glass, copper layer, and interface layer under the same temperature change. Based on this, the CTE mismatch strain between the glass and copper layer is expressed as the difference in their free thermal strains.

[0037] ;

[0038] Or, under the constant coefficient approximation, it can be converted to:

[0039] ;

[0040] Sub-model of residual effects in copper plating: The residual effects in copper plating are related to copper layer thickness, deposition rate, deposition time, electroplating current density, additive concentration, plating bath temperature, grain structure, and subsequent annealing conditions; the set of copper plating process parameters is represented as follows:

[0041] ;

[0042] in, For the thickness of the copper layer, For the deposition rate, For the deposition time, For copper plating or electroplating current density, , , These represent the concentrations of accelerator, inhibitor, and leveling agent, respectively. The temperature of the plating solution. and These are the annealing temperature and annealing time, respectively.

[0043] The residual effect of copper plating is expressed in the form of equivalent residual strain as follows:

[0044] ;

[0045] Alternatively, it can be expressed in the form of initial residual stress as follows:

[0046] ;

[0047] in, and It is an calibrable function. and The parameter to be calibrated is the calibrable function, which is an empirical-physical hybrid mapping relationship established based on copper plating tests, copper layer curvature or warpage measurements, residual stress tests, stress change data before and after annealing, or numerical inversion results.

[0048] Thermoelastic constitutive relation sub-model: CTE mismatch thermal strain and copper plating equivalent residual strain are not directly equivalent to stress, but enter the thermoelastic constitutive relation as free strain or intrinsic strain terms, and together with the total strain generated by the displacement field, they determine the actual stress field.

[0049] For the In the material-like region, under deformation and linear thermoelastic conditions, the actual stress field is expressed as:

[0050] ;

[0051] in, For the first The actual stress field within the material-like region is calculated from elastic strain and used as input for subsequent mechanical equilibrium and defect risk assessment. For the first similar materials at temperature The elastic stiffness matrix under the following conditions The total strain is generated by the displacement field. For thermal strain, This represents the equivalent residual strain corresponding to the residual effect of copper plating.

[0052] If the initial residual stress form is used, then the total stress Represented as:

[0053] ;

[0054] Mechanical equilibrium-displacement-warp model:

[0055] Substituting the obtained total stress field into the mechanical equilibrium equations, and combining the clamping constraints, support conditions, pressing force, and mechanical boundary conditions, the displacement field, stress field, and warping state of the glass substrate are solved; under deformation conditions, the strain-displacement relationship is:

[0056] ;

[0057] in, For the displacement field, the quasi-static equilibrium equations are:

[0058] ;

[0059] in, For physical force or equivalent mechanical load; The total stress field is obtained from the constitutive relation; the strain-displacement relation, constitutive relation, mechanical equilibrium equations and mechanical boundary conditions together form a set of mechanical solution equations, and the mechanical equilibrium equations are not derived from the strain-displacement relation alone.

[0060] The warpage is defined by the displacement field in the thickness direction as:

[0061] ;

[0062] in, This represents the displacement component in the thickness direction. The calculation area is for the glass substrate;

[0063] Local stress concentration and interface failure sub-model: After obtaining the total stress field and displacement field, the system calculates local stress concentration for geometric or material discontinuities, including TGV hole edges, hole wall corners, hole array gaps, copper layer edges, RDL line corners, and glass-metal interfaces; the local stress concentration factor is defined as:

[0064] ;

[0065] in, This represents the local maximum principal stress or maximum equivalent stress. Nominal stress; and The same type of stress index should be used; for the defect risk assessment of glass substrates, the maximum principal tensile stress should be used as the evaluation index for local stress concentration.

[0066] For the risk of glass-copper interface cracking or delamination, the secondary nominal stress criterion is adopted:

[0067] ;

[0068] in, For the interface normal stress, and These are two tangential shear stress components. , , These are the threshold values ​​for interface normal intensity and tangential intensity, respectively. ;when When the interface region reaches the interface damage initiation threshold, there is a risk of crack initiation or delamination.

[0069] If describing crack propagation or delamination propagation, the energy release rate criterion is used:

[0070] ;

[0071] when At that time, it was determined that the interface crack had a risk of propagation, among which Energy release rate, This is the critical fracture energy at the interface; for mixed-type interface cracks... Expressed as the sum of energy release rates of different fracture modes:

[0072] ;

[0073] The energy release rate of the type I fracture mode; The energy release rate of the type II fracture mode; Energy release rate of type III fracture mode;

[0074] The critical fracture energy at the interface can be written as a critical value related to the crack mixed mode. ,in These are mixed modal parameters.

[0075] Furthermore, the constructed training samples are represented as follows:

[0076] ;

[0077] in, This represents a set of simulation samples under multiple operating conditions. Indicates the first Numerical solution of the working conditions. Represents a sequence of historical states. Indicates action input, Indicates the next state. This indicates the defect risk label corresponding to the next state. Indicates the number of samples;

[0078] The current state is represented as:

[0079] ;

[0080] in, Represents the temperature field. Represents the current density field. Indicates the Joule heat source, Represents the stress field. Represents the displacement field. Indicates a warped state. Indicates the residual stress state of the copper plating. Indicates the local stress concentration factor. Indicators representing the risk of interface failure. Indicates the characteristics of detection observations or defects;

[0081] Action input is represented as:

[0082] ;

[0083] in, This indicates adjustable geometric movements during the design phase. Indicates thermal process actions, Indicates electrical loading action, This indicates the copper plating process. This indicates the clamping action; the action input does not include temperature field, stress field, displacement field, warpage, residual stress field, and risk label obtained from simulation or detection.

[0084] Furthermore, three strategies are adopted in the training sample generation stage: First, uniformly cover the geometry, TGV structure, copper layer thickness, temperature profile, current density, and clamping boundary within the typical process window; second, randomize the uncertain parameters of material thermal expansion coefficient, thermal conductivity, resistivity, interface strength, heat transfer coefficient, and deposition rate within a set range; third, retain some glass dimensions, TGV aperture arrays, or extreme process conditions separately as cross-condition verification samples.

[0085] Furthermore, the action-conditional world model includes:

[0086] A multimodal state encoder is used to receive multi-source data input and uniformly encode historical state sequences, current states, simulated physical fields, detection and observation images, and process parameters to obtain a latent state representation. The process is represented as:

[0087] ;

[0088] in, The parameter is A multimodal state encoder, Indicates the first The potential state representation of each process step;

[0089] Action-conditional latent dynamics model, used to predict the next latent state based on historical latent state sequences and action inputs, is represented as follows:

[0090] ;

[0091] in, The parameter is The potential dynamics model of the action conditions Indicates from the first The process steps up to the first The potential state sequence of each process step Indicates action input, Indicates the next potential state;

[0092] Multi-head decoder, used to determine the next potential state Decoding to predict the output can be represented as follows:

[0093] ;

[0094] in, The parameter is Multi-head decoder, To predict the temperature field, To predict the stress field, To predict the displacement field, To predict the warping state, To predict defect risks, the multi-head decoder includes a temperature head, a stress head, a displacement head, and a defect risk head, which output the temperature field, stress field, displacement or warping, and defect risk, respectively.

[0095] Furthermore, supervised-physical constraint joint training includes:

[0096] The monitoring loss is expressed as:

[0097] ;

[0098] in, The term represents the supervision loss. The first term constrains the error between the predicted state and the numerical simulation labeled state, while the second term constrains the error between the defect risk prediction and the defect risk label. These are the weighting coefficients. Cross-entropy loss, mean squared error, or risk ranking loss can be used.

[0099] The physical loss constraint is expressed as:

[0100] ;

[0101] in, This represents the loss of heat propagation consistency, used to constrain the temperature field to be consistent with the Joule heat source and the laws of heat diffusion; It represents the loss of mechanical equilibrium consistency and is used to constrain the stress field to satisfy the thermoelastic constitutive relation and the structural equilibrium equation; This indicates the loss due to the evolution of residual stress in the copper plating process. This represents the interface strength or interface failure constraint loss. , , , These are the corresponding weighting coefficients;

[0102] The total training loss of the model is expressed as:

[0103] ;

[0104] in, This indicates the loss of timing consistency, used to avoid unreasonable jumps between consecutive process steps; This represents the confidence assessment loss, used to guide the model output to match the prediction uncertainty with the error level. , , These are the weighting coefficients; supervised-physical-constrained training and parameter updates are completed by minimizing the total training loss of the model.

[0105] Furthermore, the training of the action-conditional world model also includes a generalization enhancement training mechanism, which includes:

[0106] The cross-condition verification strategy intentionally retains certain complete families of conditions as a verification set during training to check whether the model can predict structural and process conditions that do not appear directly in the training set.

[0107] The working condition perturbation consistency training applies perturbations to material properties, heat transfer coefficients, resistivity, deposition rate, and interface strength parameters within a reasonable range, requiring the model output temperature field, stress field, and risk level to maintain physically reasonable continuous changes; for samples close to the failure threshold, the model allows changes in the risk level, but these changes must be consistent with the direction of stress concentration, interface strength margin, or temperature gradient changes.

[0108] The confidence assessment strategy obtains the prediction confidence through deep ensemble models, Monte Carlo Dropout, quantile regression, or conformal prediction methods; for continuous output of temperature field, stress field, and warpage, the prediction interval is output; for output of crack, delamination, and copper layer anomaly risks, the risk probability and its confidence interval are output.

[0109] The active sampling mechanism selects operating conditions that need to be supplemented for simulation based on the following conditions after the initial model training is completed: operating conditions where the cross-operating condition verification error is greater than the corresponding set value, operating conditions where the model confidence is lower than the corresponding set value or the prediction interval is too wide, operating conditions where the physical residual is higher than the set value, and operating conditions that are located near the risk boundary and have a greater impact on process decisions than the corresponding set value. The action conditions of the operating conditions are returned, the simulation is repeated, supplementary samples are generated and added to the incremental training sample library, forming a closed loop of "initial simulation sample - action condition world model training - cross-operating condition verification - confidence assessment - active sampling - incremental training".

[0110] Furthermore, when predicting the next state and future multi-step states of the glass substrate through the trained action condition world model, the input action conditions include geometric parameters, thermal process parameters, electrical loading parameters, copper plating process parameters, and clamping parameters.

[0111] Based on the current state and action conditions, real-time reasoning is performed to predict the next state, and multi-step autoregressive prediction is used to predict future multi-step states.

[0112] The multi-step autoregressive prediction is expressed as follows:

[0113] ;

[0114] in, , Rolling prediction step size, Indicates the first Step-by-step prediction of state, Indicates the first Step-by-step risk prediction It represents a rolling historical state sequence composed of both the actual historical state and the preceding predicted state;

[0115] The predicted output temperature field, stress field, displacement field, warping state, and defect risk distribution are visualized.

[0116] Based on the prediction results and preset process constraints, risk thresholds and optimization objectives, suggestions for parameter correction, online control adjustment or process replanning are given. When the prediction results are in the set low-risk range, it is recommended to continue the current process. When they are in the set medium-risk range, parameter adjustment suggestions are generated. When they are in the set high-risk range, a risk warning is issued and a process replanning suggestion is proposed.

[0117] Furthermore, the method includes using the trained action condition world model as a fast predictor and combining it with a multi-objective optimization algorithm to optimize process parameters; using at least one adjustable parameter in the action conditions as an optimization variable, and using one or more of the following as optimization objectives: minimum defect risk, minimum warpage, minimum temperature gradient, minimum maximum stress or minimum local stress concentration factor, to generate a set of candidate action conditions using a multi-objective optimization algorithm; then, the world model performs fast prediction and sorting of the candidate parameter combinations, and outputs a recommended process scheme that meets the defect risk threshold, physical constraints and process constraints.

[0118] Furthermore, to achieve the above objectives, the present invention also provides a glass substrate defect prediction and process optimization system driven by a physical constraint world model, used to implement the above-mentioned glass substrate defect prediction and process optimization method driven by a physical constraint world model. The system includes:

[0119] The process data acquisition module is used to acquire process data of the glass substrate;

[0120] The mechanism modeling module is used to establish a model of the thermo-electric-mechanical relationship mechanism.

[0121] The numerical solution module is used to perform multi-condition numerical solutions based on the thermo-electric-mechanical correlation mechanism model and generate training samples;

[0122] The construction and training module is used to construct and train the action-conditional world model, which includes a multimodal state encoder, an action-conditional latent dynamics model, and a multi-head decoder. During the training process, supervision loss, physical loss constraint, temporal consistency loss, and confidence evaluation loss are introduced.

[0123] The online inference module is used to deploy the trained action-conditional world model, receive action-conditional input, and perform real-time state prediction and defect risk assessment.

[0124] The process optimization module is used to provide optimization suggestions for process parameters based on the prediction results.

[0125] The human-computer interaction module is used to receive action condition inputs and visualize the prediction results.

[0126] Compared with existing technologies, the principles and advantages of this technical solution are as follows:

[0127] 1. Defect prediction shifts from post-process inspection to pre-process assessment. Traditional inspection methods such as AOI, CT, and SEM can only detect defects after the process is completed. This technical solution, through the action condition world model, can predict future temperature fields, stress fields, and defect risks during the process parameter setting stage, thereby achieving preventive quality control and reducing scrap rates and rework costs.

[0128] 2. A unified expression of the thermo-electric-mechanical correlation mechanism for defect prediction has been established. Existing technologies typically focus on single physical fields or local coupling relationships. This technical solution unifies the correlation between current density field, Joule heat source, transient temperature field, CTE mismatch thermal strain, residual effect of copper plating, thermoelastic stress response, local stress concentration, and interface failure, forming a complete physical transmission chain from process operation conditions to defect risk. This can accurately identify the causes of various defects such as cracks, warpage, delamination, copper layer anomalies, and interface failure.

[0129] 3. Reduce numerical simulation time from hours to milliseconds. For complex glass substrate systems containing TGV via arrays, copper layers, interface layers, and clamping boundaries, a single high-fidelity numerical simulation may take several hours; this invention trains the world model offline, and the online inference stage can complete state prediction in just milliseconds, supporting rapid scanning of large-scale operating conditions and interactive process optimization.

[0130] 4. Physical constraint loss ensures that the model predictions conform to physical laws. Ordinary neural networks are prone to unreasonable predictions such as discontinuous temperature fields and shifted stress concentration locations when there are insufficient training samples or distribution biases. This technical solution embeds the thermo-electric-mechanical correlation mechanism into the model training through heat propagation consistency loss, mechanical equilibrium consistency loss, residual stress evolution loss of copper plating, and interface failure constraint loss, which significantly improves the physical reliability and extrapolation ability of the prediction results.

[0131] 5. The generalization enhancement training mechanism improves the predictive ability for unseen working conditions. Through cross-working condition verification, consistent training under working condition disturbances, confidence assessment, and active sampling mechanism, this technical solution can effectively reduce the risk of overfitting the action condition world model to a limited number of simulation samples, enabling it to provide reliable predictions for unseen working conditions such as glass size, TGV structure, temperature profile, and clamping method that are not directly present in the training set.

[0132] 6. Closed-loop iterative mechanism continuously optimizes model performance. The closed-loop mechanism of "initial simulation samples - world model training - cross-condition verification - confidence assessment - active sampling - incremental training" enables the system to automatically identify weak areas in prediction and supplement simulation samples, thereby achieving continuous iterative improvement of model performance and adapting to the prediction needs of new processes, new materials and new structures. Attached Figure Description

[0133] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0134] Figure 1 Flowchart for modeling the thermo-electric-mechanical correlation mechanism, numerical solution under multiple operating conditions, and sample generation;

[0135] Figure 2 This is a flowchart of the offline training and online inference process for the action-conditional world model. Detailed Implementation

[0136] The present invention will be further described below with reference to specific embodiments:

[0137] The physical constraint world model-driven method for glass substrate defect prediction and process optimization described in this embodiment includes the following steps:

[0138] S1. Obtain the process data of the glass substrate;

[0139] In this step, the process data includes four categories:

[0140] (1) Geometric data, including the length, width and thickness of the glass substrate, the diameter, spacing, depth and aspect ratio of the TGV aperture, the aperture array arrangement, the cone angle of the aperture wall, the structure of the RDL region, and the thickness of the copper or metal layer.

[0141] (2) Material property data, including the coefficient of thermal expansion, thermal conductivity, specific heat capacity, density, resistivity, elastic modulus, Poisson's ratio, interfacial bonding strength and fracture-related parameters of glass, copper layer and interface layer.

[0142] (3) Process boundary conditions, including temperature curve, heating rate, cooling rate, holding temperature, holding time, current density, voltage or power, energizing time, clamping method, support position, pressing force, pressure boundary and cooling boundary.

[0143] (4) Copper plating process parameters, including target copper layer thickness, deposition time, deposition rate, ternary agent ratio, plating solution temperature, flow rate or stirring conditions, as well as heat treatment temperature and annealing time.

[0144] S2, such as Figure 1 As shown, a thermo-electric-mechanical correlation mechanism model is established based on the process data; the thermo-electric-mechanical correlation mechanism model is used to describe the correlation between the current density field and Joule heat source, transient heat propagation, CTE mismatch thermal strain, residual effect of copper plating, thermoelastic stress response, local stress concentration, and interface crack or delamination failure.

[0145] Specifically, the thermo-electric-mechanical correlation mechanism model includes the electric field-Joule heat source sub-model, the transient heat propagation sub-model, the CTE mismatch thermal strain sub-model, the residual effect of copper plating sub-model, the thermoelastic constitutive relation sub-model, the mechanical equilibrium-displacement-warping sub-model, and the local stress concentration and interface failure sub-model.

[0146] The electric field-Joule heat source sub-model is constructed as follows:

[0147] First, solve for the current density. :

[0148] ;

[0149] For electric field strength, For electric potential, The conductivity varies with temperature; Represents the electric potential Gradient calculation;

[0150] The governing equation for electric potential is:

[0151] ;

[0152] For spatial differential operators;

[0153] For a Joule heat source within a conductive region, it is represented as:

[0154] ;

[0155] When expressed in terms of resistivity and current density, the Joule heat source is written as:

[0156] ;

[0157] Resistivity;

[0158] The transient heat propagation sub-model is expressed as follows:

[0159] ;

[0160] For density, For specific heat capacity, Thermal conductivity, For temperature, As a Joule heat source, Other internal heat sources besides Joule heating;

[0161] CTE mismatch thermal strain quantum model: for the first For this type of material, the thermal strain is expressed as:

[0162] ;

[0163] in, For the first The coefficient of thermal expansion of similar materials, For reference temperature, It is a unit tensor; when the coefficient of thermal expansion is approximately constant within the temperature range, it simplifies to:

[0164] ;

[0165] First calculate the first The thermal strain of these materials is used to describe the free expansion trends of the glass, copper layer, and interface layer under the same temperature change. Based on this, the CTE mismatch strain between the glass and copper layer is expressed as the difference in their free thermal strains.

[0166] ;

[0167] Or, under the constant coefficient approximation, it can be converted to:

[0168] ;

[0169] Sub-model of residual effects in copper plating: The residual effects in copper plating are related to copper layer thickness, deposition rate, deposition time, electroplating current density, additive concentration, plating bath temperature, grain structure, and subsequent annealing conditions; the set of copper plating process parameters is represented as follows:

[0170] ;

[0171] in, For the thickness of the copper layer, For the deposition rate, For the deposition time, For copper plating or electroplating current density, , , These represent the concentrations of accelerator, inhibitor, and leveling agent, respectively. The temperature of the plating solution. and These are the annealing temperature and annealing time, respectively.

[0172] The residual effect of copper plating is expressed in the form of equivalent residual strain as follows:

[0173] ;

[0174] Alternatively, it can be expressed in the form of initial residual stress as follows:

[0175] ;

[0176] in, and It is an calibrable function. and The parameter to be calibrated is the calibrable function, which is an empirical-physical hybrid mapping relationship established based on copper plating tests, copper layer curvature or warpage measurements, residual stress tests, stress change data before and after annealing, or numerical inversion results.

[0177] Thermoelastic constitutive relation sub-model: CTE mismatch thermal strain and copper plating equivalent residual strain are not directly equivalent to stress, but enter the thermoelastic constitutive relation as free strain or intrinsic strain terms, and together with the total strain generated by the displacement field, they determine the actual stress field.

[0178] For the In the material-like region, under deformation and linear thermoelastic conditions, the actual stress field is expressed as:

[0179] ;

[0180] in, For the first The actual stress field within the material-like region is calculated from elastic strain and used as input for subsequent mechanical equilibrium and defect risk assessment. For the first similar materials at temperature The elastic stiffness matrix under the following conditions The total strain is generated by the displacement field. For thermal strain, This represents the equivalent residual strain corresponding to the residual effect of copper plating.

[0181] If the initial residual stress form is used, then the total stress Represented as:

[0182] ;

[0183] Mechanical equilibrium-displacement-warp model:

[0184] Substituting the obtained total stress field into the mechanical equilibrium equations, and combining the clamping constraints, support conditions, pressing force, and mechanical boundary conditions, the displacement field, stress field, and warping state of the glass substrate are solved; under deformation conditions, the strain-displacement relationship is:

[0185] ;

[0186] in, For the displacement field, the quasi-static equilibrium equations are:

[0187] ;

[0188] in, For physical force or equivalent mechanical load; The total stress field is obtained from the constitutive relation; the strain-displacement relation, constitutive relation, mechanical equilibrium equations and mechanical boundary conditions together form a set of mechanical solution equations, and the mechanical equilibrium equations are not derived from the strain-displacement relation alone.

[0189] The warpage is defined by the displacement field in the thickness direction as:

[0190] ;

[0191] in, This represents the displacement component in the thickness direction. The calculation area is for the glass substrate;

[0192] Local stress concentration and interface failure sub-model: After obtaining the total stress field and displacement field, the system calculates local stress concentration for geometric or material discontinuities, including TGV hole edges, hole wall corners, hole array gaps, copper layer edges, RDL line corners, and glass-metal interfaces; the local stress concentration factor is defined as:

[0193] ;

[0194] in, This represents the local maximum principal stress or maximum equivalent stress. Nominal stress; and The same type of stress index should be used; for the defect risk assessment of glass substrates, the maximum principal tensile stress should be used as the evaluation index for local stress concentration.

[0195] For the risk of glass-copper interface cracking or delamination, the secondary nominal stress criterion is adopted:

[0196] ;

[0197] in, For the interface normal stress, and These are two tangential shear stress components. , , These are the threshold values ​​for interface normal intensity and tangential intensity, respectively. ;when When the interface region reaches the interface damage initiation threshold, there is a risk of crack initiation or delamination.

[0198] If describing crack propagation or delamination propagation, the energy release rate criterion is used:

[0199] ;

[0200] when At that time, it was determined that the interface crack had a risk of propagation, among which Energy release rate, This is the critical fracture energy at the interface; for mixed-type interface cracks... Expressed as the sum of energy release rates of different fracture modes:

[0201] ;

[0202] The energy release rate of the type I fracture mode; The energy release rate of the type II fracture mode; Energy release rate of type III fracture mode;

[0203] The critical fracture energy at the interface can be written as a critical value related to the crack mixed mode. ,in These are mixed modal parameters.

[0204] S3, such as Figure 1 As shown, the process data is used as the variable operating condition input, and multi-condition numerical solutions are performed based on the thermo-electric-mechanical correlation mechanism model to obtain numerical solution results; and training samples are constructed based on the numerical solution results.

[0205] Specifically, after obtaining the thermo-electric-mechanical correlation mechanism model, a multi-condition simulation design matrix is ​​constructed using geometric structure data, material property data, process boundary conditions, and copper plating process parameters as variable operating condition inputs. Numerical solutions for the thermo-electric-mechanical correlation are then performed for different combinations of operating conditions. For each set of operating conditions, a geometric model is first established based on the glass size, TGV aperture, hole spacing, hole depth, hole array arrangement, and copper layer thickness. Mesh refinement is then applied to the TGV hole edges, hole wall corners, hole array edges, copper layer edges, and the glass-metal interface region. Subsequently, the established thermo-electric-mechanical correlation mechanism model is spatially discretized and solved using the finite element method, finite volume method, or finite difference method. Based on the solution results, assessments of the temperature field, stress field, displacement field, warpage state, residual stress state, and interface failure risk are conducted.

[0206] The numerical simulation output includes temperature field, current density field, Joule heat source distribution, stress field, displacement field, warpage state, interface stress distribution, and interface strength margin. Based on preset physical thresholds and defect criteria, the system labels defect risks in high-temperature gradient regions, high-stress regions, residual stress concentration regions, warpage anomaly regions, and interface failure regions, generating tags for crack risk, delamination risk, warpage risk, copper layer anomaly risk, and interface failure risk.

[0207] To improve the coverage of the initial sample library, the system employs three strategies during the sample generation phase: First, it uniformly covers the geometry, TGV structure, copper layer thickness, temperature profile, current density, and clamping boundaries within typical process windows; second, it randomizes uncertain parameters such as the material's thermal expansion coefficient, thermal conductivity, resistivity, interface strength, heat transfer coefficient, and deposition rate within a reasonable range; and third, it separately retains some glass dimensions, TGV aperture arrays, or extreme process conditions as cross-condition verification samples for subsequent checks on the model's generalization ability to conditions not directly present in the training set.

[0208] S4, such as Figure 2 As shown, an action-conditional world model is constructed, and the model is trained offline using the training samples. During the training process, supervision loss, physical loss constraint, temporal consistency loss, and confidence evaluation loss are introduced. The model parameters are updated through joint training of supervision and physical constraints. The physical loss constraint is constructed based on the thermo-electric-mechanical correlation mechanism model.

[0209] Specifically, in this step, the action-conditional world model constructed includes:

[0210] A multimodal state encoder is used to receive multi-source data input and uniformly encode historical state sequences, current states, simulated physical fields, detection and observation images, and process parameters to obtain a latent state representation. The process is represented as:

[0211] ;

[0212] in, The parameter is A multimodal state encoder, Indicates the first The potential state representation of each process step;

[0213] Action-conditional latent dynamics model, used to predict the next latent state based on historical latent state sequences and action inputs, is represented as follows:

[0214] ;

[0215] in, The parameter is The potential dynamics model of the action conditions Indicates from the first The process steps up to the first The potential state sequence of each process step Indicates action input, Indicates the next potential state;

[0216] Multi-head decoder, used to determine the next potential state Decoding to predict the output can be represented as follows:

[0217] ;

[0218] in, The parameter is Multi-head decoder, To predict the temperature field, To predict the stress field, To predict the displacement field, To predict the warping state, To predict defect risks, the multi-head decoder includes a temperature head, a stress head, a displacement head, and a defect risk head, which output the temperature field, stress field, displacement or warping, and defect risk, respectively.

[0219] Supervised-physical-constraint joint training includes:

[0220] The monitoring loss is expressed as:

[0221] ;

[0222] in, The term represents the supervision loss. The first term constrains the error between the predicted state and the numerical simulation labeled state, while the second term constrains the error between the defect risk prediction and the defect risk label. These are the weighting coefficients. Cross-entropy loss, mean squared error, or risk ranking loss can be used.

[0223] The physical loss constraint is expressed as:

[0224] ;

[0225] in, This represents the loss of heat propagation consistency, used to constrain the temperature field to be consistent with the Joule heat source and the laws of heat diffusion; It represents the loss of mechanical equilibrium consistency and is used to constrain the stress field to satisfy the thermoelastic constitutive relation and the structural equilibrium equation; This indicates the loss due to the evolution of residual stress in the copper plating process. This represents the interface strength or interface failure constraint loss. , , , These are the corresponding weighting coefficients;

[0226] The total training loss of the model is expressed as:

[0227] ;

[0228] in, This indicates the loss of timing consistency, used to avoid unreasonable jumps between consecutive process steps; This represents the confidence assessment loss, used to guide the model output to match the prediction uncertainty with the error level. , , These are the weighting coefficients; supervised-physical-constrained training and parameter updates are completed by minimizing the total training loss of the model.

[0229] The training of the action-conditional world model also includes a generalization enhancement training mechanism, which includes:

[0230] The cross-condition validation strategy (to improve the reliability of the model for unseen conditions) deliberately retains certain complete families of conditions as a validation set during training to check whether the model can predict structural and process conditions that do not appear directly in the training set.

[0231] Operating condition perturbation consistency training (to enhance model stability) involves applying perturbations to material properties, heat transfer coefficients, resistivity, deposition rates, and interface strength parameters within a reasonable range. The model output temperature field, stress field, and risk level are required to maintain physically reasonable and continuous changes. For samples close to the failure threshold, the model allows changes in the risk level, but these changes must be consistent with the direction of stress concentration, interface strength margin, or temperature gradient changes.

[0232] The confidence assessment strategy (to determine whether the model prediction is reliable) obtains the prediction confidence through deep ensemble model, Monte Carlo Dropout, quantile regression or conformal prediction methods; for continuous output of temperature field, stress field and warpage, the prediction interval is output; for output of crack, delamination and copper layer anomaly risks, the risk probability and its confidence interval are output.

[0233] The active sampling mechanism selects operating conditions that need to be supplemented for simulation based on the following conditions after the initial model training is completed: operating conditions where the cross-operating condition verification error is greater than the corresponding set value, operating conditions where the model confidence is lower than the corresponding set value or the prediction interval is too wide, operating conditions where the physical residual is higher than the set value, and operating conditions that are located near the risk boundary and have a greater impact on process decisions than the corresponding set value. The action conditions of the operating conditions are returned, the simulation is repeated, supplementary samples are generated and added to the incremental training sample library, forming a closed loop of "initial simulation sample - action condition world model training - cross-operating condition verification - confidence assessment - active sampling - incremental training".

[0234] S5. Predict the next state and future multi-step states of the glass substrate using the trained action condition world model, and provide process parameter optimization suggestions based on the prediction results.

[0235] In this step, when predicting the next state and future multi-step states of the glass substrate using the trained action condition world model, the input action conditions include geometric parameters, thermal process parameters, electrical loading parameters, copper plating process parameters, and clamping parameters.

[0236] Based on the current state and action conditions, real-time reasoning is performed to predict the next state, and multi-step autoregressive prediction is used to predict future multi-step states.

[0237] The multi-step autoregressive prediction is expressed as follows:

[0238] ;

[0239] in, , Rolling prediction step size, Indicates the first Step-by-step prediction of state, Indicates the first Step-by-step risk prediction It represents a rolling historical state sequence composed of both the actual historical state and the preceding predicted state;

[0240] The predicted output temperature field, stress field, displacement field, warping state, and defect risk distribution are visualized.

[0241] Based on the prediction results and preset process constraints, risk thresholds, and optimization objectives, suggestions for parameter correction, online control adjustment, or process replanning are provided. When the prediction results are in the set low-risk range, it is recommended to continue the current process; when they are in the set medium-risk range, parameter adjustment suggestions are generated; and when they are in the set high-risk range, a risk warning is issued and process replanning suggestions are proposed. For example, reducing the rate of temperature change, adjusting the copper layer thickness, shortening the energizing time, and optimizing the ternary agent ratio can reduce the risk of defect formation.

[0242] Furthermore, the trained action condition world model can be used as a fast predictor, combined with a multi-objective optimization algorithm to optimize process parameters. At least one adjustable parameter in the action condition is used as the optimization variable, and one or more of the following are used as optimization objectives: minimum defect risk, minimum warpage, minimum temperature gradient, minimum maximum stress or minimum local stress concentration factor. A set of candidate action conditions is generated using a multi-objective optimization algorithm. Then, the world model quickly predicts and sorts the candidate parameter combinations, and outputs a recommended process scheme that meets the defect risk threshold, physical constraints and process constraints.

[0243] In addition, this embodiment also includes a glass substrate defect prediction and process optimization system driven by a physical constraint world model, used to implement the above-mentioned glass substrate defect prediction and process optimization method driven by a physical constraint world model. The system includes:

[0244] The process data acquisition module is used to acquire process data of the glass substrate;

[0245] The mechanism modeling module is used to establish a model of the thermo-electric-mechanical relationship mechanism.

[0246] The numerical solution module is used to perform multi-condition numerical solutions based on the thermo-electric-mechanical correlation mechanism model and generate training samples;

[0247] The construction and training module is used to construct and train the action-conditional world model, which includes a multimodal state encoder, an action-conditional latent dynamics model, and a multi-head decoder. During the training process, supervision loss, physical loss constraint, temporal consistency loss, and confidence evaluation loss are introduced.

[0248] The online inference module is used to deploy the trained action-conditional world model, receive action-conditional input, and perform real-time state prediction and defect risk assessment.

[0249] The process optimization module is used to provide optimization suggestions for process parameters based on the prediction results.

[0250] The human-computer interaction module is used to receive action condition inputs and visualize the prediction results.

[0251] The above-described embodiments are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Therefore, any changes made in accordance with the shape and principle of the present invention should be covered within the protection scope of the present invention.

Claims

1. A method for defect prediction and process optimization of glass substrates driven by a physically constrained world model, characterized in that, include: Obtain process data for the glass substrate; A thermo-electric-mechanical correlation mechanism model was established based on the aforementioned process data; The thermo-electric-mechanical correlation mechanism model is used to describe the correlation between the current density field and Joule heat source, transient heat propagation, CTE mismatch thermal strain, residual effect of copper plating, thermoelastic stress response, local stress concentration, and interface crack or delamination failure. Using the process data as variable operating condition input, multi-condition numerical solutions are performed based on the thermo-electric-mechanical correlation mechanism model to obtain numerical solution results; Training samples are constructed based on the numerical solution results; An action-conditional world model is constructed, and the model is trained offline using the training samples. During the training process, supervised loss, physical loss constraint, temporal consistency loss, and confidence evaluation loss are introduced. The model parameters are updated through joint training of supervised and physical constraints. The physical loss constraint is constructed based on the thermo-electric-mechanical correlation mechanism model. The trained action conditional world model predicts the next state and future multiple states of the glass substrate, and provides optimization suggestions for process parameters based on the prediction results.

2. The method for glass substrate defect prediction and process optimization driven by a physical constraint world model according to claim 1, characterized in that, The thermo-electric-mechanical correlation mechanism model includes the electric field-Joule heat source sub-model, transient heat propagation sub-model, CTE mismatch thermal strain sub-model, copper plating residual effect sub-model, thermoelastic constitutive relation sub-model, mechanical equilibrium-displacement-warping sub-model, and local stress concentration and interface failure sub-model. The electric field-Joule heat source sub-model is constructed as follows: First, solve for the current density. : ； For electric field strength, For electric potential, The conductivity varies with temperature; Represents the electric potential Gradient calculation; The governing equation for electric potential is: ； For spatial differential operators; For a Joule heat source within a conductive region, it is represented as: ； When expressed in terms of resistivity and current density, the Joule heat source is written as: ； Resistivity; The transient heat propagation sub-model is expressed as follows: ； For density, For specific heat capacity, Thermal conductivity, For temperature, As a Joule heat source, Other internal heat sources besides Joule heating; CTE mismatch thermal strain quantum model: for the first For this type of material, the thermal strain is expressed as: ； in, For the first The coefficient of thermal expansion of similar materials, For reference temperature, It is a unit tensor; when the coefficient of thermal expansion is approximately constant within the temperature range, it simplifies to: ； First calculate the first The thermal strain of these materials is used to describe the free expansion trends of the glass, copper layer, and interface layer under the same temperature change. Based on this, the CTE mismatch strain between the glass and copper layer is expressed as the difference in their free thermal strains. ； Or, under the constant coefficient approximation, it can be converted to: ； Sub-model of residual effects in copper plating: The residual effects in copper plating are related to copper layer thickness, deposition rate, deposition time, electroplating current density, additive concentration, plating bath temperature, grain structure, and subsequent annealing conditions; the set of copper plating process parameters is represented as follows: ； in, For the thickness of the copper layer, For the deposition rate, For the deposition time, For copper plating or electroplating current density, , , These represent the concentrations of accelerator, inhibitor, and leveling agent, respectively. The temperature of the plating solution. and These are the annealing temperature and annealing time, respectively. The residual effect of copper plating is expressed in the form of equivalent residual strain as follows: ； Alternatively, it can be expressed in the form of initial residual stress as follows: ； in, and It is an calibrable function. and The parameter to be calibrated is the calibrable function, which is an empirical-physical hybrid mapping relationship established based on copper plating tests, copper layer curvature or warpage measurements, residual stress tests, stress change data before and after annealing, or numerical inversion results. Thermoelastic constitutive relation sub-model: CTE mismatch thermal strain and copper plating equivalent residual strain are not directly equivalent to stress, but enter the thermoelastic constitutive relation as free strain or intrinsic strain terms, and together with the total strain generated by the displacement field, they determine the actual stress field. For the In the material-like region, under deformation and linear thermoelastic conditions, the actual stress field is expressed as: ； in, For the first The actual stress field within the material-like region is calculated from elastic strain and used as input for subsequent mechanical equilibrium and defect risk assessment. For the first similar materials at temperature The elastic stiffness matrix under the following conditions The total strain is generated by the displacement field. For thermal strain, This represents the equivalent residual strain corresponding to the residual effect of copper plating. If the initial residual stress form is used, then the total stress Represented as: ； Mechanical equilibrium-displacement-warp model: Substituting the obtained total stress field into the mechanical equilibrium equations, and combining the clamping constraints, support conditions, pressing force, and mechanical boundary conditions, the displacement field, stress field, and warping state of the glass substrate are solved; under deformation conditions, the strain-displacement relationship is: ； in, For the displacement field, the quasi-static equilibrium equations are: ； in, For physical force or equivalent mechanical load; The total stress field is obtained from the constitutive relation; the strain-displacement relation, constitutive relation, mechanical equilibrium equations and mechanical boundary conditions together form a set of mechanical solution equations, and the mechanical equilibrium equations are not derived from the strain-displacement relation alone. The warpage is defined by the displacement field in the thickness direction as: ； in, This represents the displacement component in the thickness direction. The calculation area is for the glass substrate; Local stress concentration and interface failure sub-model: After obtaining the total stress field and displacement field, the system calculates local stress concentration for geometric or material discontinuities, including TGV hole edges, hole wall corners, hole array gaps, copper layer edges, RDL line corners, and glass-metal interfaces; the local stress concentration factor is defined as: ； in, This represents the local maximum principal stress or maximum equivalent stress. Nominal stress; and The same type of stress index should be used; for the defect risk assessment of glass substrates, the maximum principal tensile stress should be used as the evaluation index for local stress concentration. For the risk of glass-copper interface cracking or delamination, the secondary nominal stress criterion is adopted: ； in, For the interface normal stress, and These are two tangential shear stress components. , , These are the threshold values ​​for interface normal intensity and tangential intensity, respectively. ;when When the interface region reaches the interface damage initiation threshold, there is a risk of crack initiation or delamination. If describing crack propagation or delamination propagation, the energy release rate criterion is used: ； when At that time, it was determined that the interface crack had a risk of propagation, among which Energy release rate, This is the critical fracture energy at the interface; for mixed-type interface cracks... Expressed as the sum of energy release rates of different fracture modes: ； The energy release rate of the type I fracture mode; The energy release rate of the type II fracture mode; Energy release rate of type III fracture mode; The critical fracture energy at the interface can be written as a critical value related to the crack mixed mode. ,in These are mixed modal parameters.

3. The method for glass substrate defect prediction and process optimization driven by a physical constraint world model according to claim 2, characterized in that, The constructed training samples are represented as follows: ； in, This represents a set of simulation samples under multiple operating conditions. Indicates the first Numerical solution of the working conditions. Represents a sequence of historical states. Indicates action input, Indicates the next state. This indicates the defect risk label corresponding to the next state. Indicates the number of samples; The current state is represented as: ； in, Represents the temperature field. Represents the current density field. Indicates the Joule heat source, Represents the stress field. Represents the displacement field. Indicates a warped state. Indicates the residual stress state of the copper plating. Indicates the local stress concentration factor. Indicators representing the risk of interface failure. Indicates the characteristics of detection observations or defects; Action input is represented as: ； in, This indicates adjustable geometric movements during the design phase. Indicates thermal process actions, Indicates electrical loading action, This indicates the copper plating process. This indicates the clamping action; the action input does not include temperature field, stress field, displacement field, warpage, residual stress field, and risk label obtained from simulation or detection.

4. The method for glass substrate defect prediction and process optimization driven by a physical constraint world model according to claim 3, characterized in that, Three strategies were adopted during the training sample generation phase: First, uniform coverage was applied to the geometry, TGV structure, copper layer thickness, temperature profile, current density, and clamping boundary within a typical process window; second, the uncertain parameters of material thermal expansion coefficient, thermal conductivity, resistivity, interface strength, heat transfer coefficient, and deposition rate were randomized within a set range; and third, some glass dimensions, TGV aperture arrays, or extreme process conditions were separately retained as cross-condition verification samples.

5. The method for glass substrate defect prediction and process optimization driven by a physical constraint world model according to claim 3, characterized in that, The action-conditional world model includes: A multimodal state encoder is used to receive multi-source data input and uniformly encode historical state sequences, current states, simulated physical fields, detection and observation images, and process parameters to obtain a latent state representation. The process is represented as: ； in, The parameter is A multimodal state encoder, Indicates the first The potential state representation of each process step; Action-conditional latent dynamics model, used to predict the next latent state based on historical latent state sequences and action inputs, is represented as follows: ； in, The parameter is The potential dynamics model of the action conditions Indicates from the first The process steps up to the first The potential state sequence of each process step Indicates action input, Indicates the next potential state; Multi-head decoder, used to determine the next potential state Decoding to predict the output can be represented as follows: ； in, The parameter is Multi-head decoder, To predict the temperature field, To predict the stress field, To predict the displacement field, To predict the warping state, To predict defect risks, the multi-head decoder includes a temperature head, a stress head, a displacement head, and a defect risk head, which output the temperature field, stress field, displacement or warping, and defect risk, respectively.

6. The method for glass substrate defect prediction and process optimization driven by a physical constraint world model according to claim 5, characterized in that, Supervised-physical-constraint joint training includes: The monitoring loss is expressed as: ； in, The term represents the supervision loss. The first term constrains the error between the predicted state and the numerical simulation labeled state, while the second term constrains the error between the defect risk prediction and the defect risk label. These are the weighting coefficients. Cross-entropy loss, mean squared error, or risk ranking loss can be used. The physical loss constraint is expressed as: ； in, This represents the loss of heat propagation consistency, used to constrain the temperature field to be consistent with the Joule heat source and the laws of heat diffusion; It represents the loss of mechanical equilibrium consistency and is used to constrain the stress field to satisfy the thermoelastic constitutive relation and the structural equilibrium equation; This indicates the loss due to the evolution of residual stress in the copper plating process. This represents the interface strength or interface failure constraint loss. , , , These are the corresponding weighting coefficients; The total training loss of the model is expressed as: ； in, This indicates the loss of timing consistency, used to avoid unreasonable jumps between consecutive process steps; This represents the confidence assessment loss, used to guide the model output to match the prediction uncertainty with the error level. , , These are the weighting coefficients; supervised-physical-constrained training and parameter updates are completed by minimizing the total training loss of the model.

7. The method for glass substrate defect prediction and process optimization driven by a physical constraint world model according to claim 6, characterized in that, The training of the action-conditional world model also includes a generalization enhancement training mechanism, which includes: The cross-condition verification strategy intentionally retains certain complete families of conditions as a verification set during training to check whether the model can predict structural and process conditions that do not appear directly in the training set. The working condition perturbation consistency training applies perturbations to material properties, heat transfer coefficients, resistivity, deposition rate, and interface strength parameters within a reasonable range, requiring the model output temperature field, stress field, and risk level to maintain physically reasonable continuous changes; for samples close to the failure threshold, the model allows changes in the risk level, but these changes must be consistent with the direction of stress concentration, interface strength margin, or temperature gradient changes. The confidence assessment strategy obtains the prediction confidence through deep ensemble models, Monte Carlo Dropout, quantile regression, or conformal prediction methods; for continuous output of temperature field, stress field, and warpage, the prediction interval is output; for output of crack, delamination, and copper layer anomaly risks, the risk probability and its confidence interval are output. The active sampling mechanism selects operating conditions that need to be supplemented for simulation based on the following conditions after the initial model training is completed: operating conditions where the cross-operating condition verification error is greater than the corresponding set value, operating conditions where the model confidence is lower than the corresponding set value or the prediction interval is too wide, operating conditions where the physical residual is higher than the set value, and operating conditions that are located near the risk boundary and have a greater impact on process decisions than the corresponding set value. The action conditions of the operating conditions are returned, the simulation is repeated, supplementary samples are generated and added to the incremental training sample library, forming a closed loop of "initial simulation sample - action condition world model training - cross-operating condition verification - confidence assessment - active sampling - incremental training".

8. The method for glass substrate defect prediction and process optimization driven by a physical constraint world model according to claim 7, characterized in that, When predicting the next state and future multi-step states of a glass substrate using the trained action condition world model, the input action conditions include geometric parameters, thermal process parameters, electrical loading parameters, copper plating process parameters, and clamping parameters. Based on the current state and action conditions, real-time reasoning is performed to predict the next state, and multi-step autoregressive prediction is used to predict future multi-step states. The multi-step autoregressive prediction is expressed as follows: ； in, , Rolling prediction step size, Indicates the first Step-by-step prediction of state, Indicates the first Step-by-step risk prediction It represents a rolling historical state sequence composed of both the actual historical state and the preceding predicted state; The predicted output temperature field, stress field, displacement field, warping state, and defect risk distribution are visualized. Based on the prediction results and preset process constraints, risk thresholds and optimization objectives, suggestions for parameter correction, online control adjustment or process replanning are given. When the prediction results are in the set low-risk range, it is recommended to continue the current process. When they are in the set medium-risk range, parameter adjustment suggestions are generated. When they are in the set high-risk range, a risk warning is issued and a process replanning suggestion is proposed.

9. The method for glass substrate defect prediction and process optimization driven by a physical constraint world model according to claim 8, characterized in that, The method also includes using the trained action condition world model as a fast predictor, combined with a multi-objective optimization algorithm to optimize process parameters; using at least one adjustable parameter in the action condition as an optimization variable, and using one or more of the following as optimization objectives: minimum defect risk, minimum warpage, minimum temperature gradient, minimum maximum stress or minimum local stress concentration factor, to generate a set of candidate action conditions using a multi-objective optimization algorithm; and then using the world model to quickly predict and sort the candidate parameter combinations, outputting a recommended process scheme that meets the defect risk threshold, physical constraints and process constraints.

10. A glass substrate defect prediction and process optimization system driven by a physical constraint world model, used to implement the glass substrate defect prediction and process optimization method driven by a physical constraint world model as described in any one of claims 1 to 9, characterized in that it comprises: The process data acquisition module is used to acquire process data of the glass substrate; The mechanism modeling module is used to establish a model of the thermo-electric-mechanical relationship mechanism. The numerical solution module is used to perform multi-condition numerical solutions based on the thermo-electric-mechanical correlation mechanism model and generate training samples; The construction and training module is used to construct and train the action-conditional world model, which includes a multimodal state encoder, an action-conditional latent dynamics model, and a multi-head decoder. During the training process, supervision loss, physical loss constraint, temporal consistency loss, and confidence evaluation loss are introduced. The online inference module is used to deploy the trained action-conditional world model, receive action-conditional input, and perform real-time state prediction and defect risk assessment. The process optimization module is used to provide optimization suggestions for process parameters based on the prediction results. The human-computer interaction module is used to receive action condition inputs and visualize the prediction results.

Images