A method for constructing a meta-model of the evolution of micro / nanostructure fabrication interfaces based on a physical information neural network constrained by Level-set equations.
By embedding Level-set equation constraints and data constraints into a physical information neural network, a main-auxiliary neural network coupled architecture is constructed, which solves the problem of high-precision prediction of cross-process parameters in the fabrication of micro and nano structures in the existing technology, and realizes efficient and accurate profile prediction and rate inversion of complex material parameters.
Patent Information
- Application Number
- CN202611621359.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-10-20
- Publication Date
- 2026-11-17
AI Technical Summary
Existing physical information neural networks struggle to achieve high generalization ability and rapid, high-precision prediction across process parameters when dealing with interface evolution problems involving complex and unknown material parameters, especially lacking unified modeling capabilities in micro and nanostructure fabrication.
A meta-model of the micro/nano structure processing interface evolution is constructed based on a physical information neural network constrained by Level-set equations. By embedding partial differential equations describing interface evolution, geometric property constraints, and data constraints into the loss function, a parameterized meta-model is constructed. A coupled architecture of main neural network and auxiliary neural network is adopted to achieve fast and high-precision profile prediction and automatic inversion of complex anisotropic rate fields for untrained process parameter combinations.
It achieves high-precision prediction of combinations of process parameters that have not been trained, improves the training efficiency and generalization of the model on small sample data, can adaptively handle the normal rate in the etching and deposition process, and is suitable for a variety of surface processing problems.
Smart Images

Figure CN123120157A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of micro-nano fabrication technology, specifically relating to a method for constructing a meta-model of the micro-nano structure fabrication interface based on a physical information neural network constrained by Level-set equations. Background Technology
[0002] In semiconductor manufacturing, microelectromechanical systems (MEMS), and other advanced materials processing fields, precise etching, deposition, or epitaxial growth of material surfaces are crucial steps in fabricating complex micro / nano structures. Currently, numerical simulations of surface profiles are mainly divided into two categories: atomic-scale methods and continuum methods. While commonly used atomic-scale methods can address microscopic mechanisms, their computational complexity explodes with increasing simulation scale and heavily relies on microscopic rate parameters that are difficult to calibrate precisely, making them unsuitable for rapid simulation and evaluation of macroscopic devices. Continuum methods treat the material surface as a continuously evolving interface, describing its motion through partial differential equations. Among these, the level-set method has become mainstream due to its ability to naturally handle topological changes and its high computational accuracy.
[0003] However, traditional level-set methods rely on Eulerian meshes arranged within the computational domain and employ finite difference and finite volume methods for solving the problem. To obtain high-precision results, dense meshes need to be arranged in three-dimensional or even higher-dimensional space, leading to an exponential increase in computational and memory requirements. Simultaneously, to ensure numerical stability, the level-set function should theoretically always remain a signed distance function, satisfying the characteristics of a signed distance function. This requires frequent re-initialization operations, which in itself is a partial differential equation problem and may introduce errors in interface position. Traditional methods typically perform calculations for a single, specific process condition; for different combinations of process parameters, the rate of change needs to be remeasured and re-simulated, lacking the ability to generalize across process parameters.
[0004] Physical Information Neural Networks (PINNs) are a novel method for solving partial differential equations that has emerged in recent years. They embed the residuals, initial conditions, and boundary conditions of the governing equations as penalty terms into the loss function of the neural network, allowing the network to automatically satisfy physical laws during training. However, existing PINN applications still have limitations in handling surface processing problems, such as difficulty in dealing with complex anisotropic inversions and insufficient decoupling of physical features. Therefore, how to further extend Physical Information Neural Networks to uniformly and efficiently handle interface evolution problems with complex unknown material parameters, and to construct a parameterized meta-model with high generalization ability, is a key challenge currently facing the field. Summary of the Invention
[0005] Objective: To address the limitation of existing physical information neural networks (PINs) in handling interface evolution with complex and unknown material parameters, this invention proposes a method for constructing a meta-model of interface evolution in micro / nanostructure processing based on a PIN constrained by Level-set equations. This method embeds partial differential equation constraints, geometric property constraints, and data constraints describing interface evolution into a loss function, constructing a parameterized meta-model. This meta-model can perform rapid and high-precision contour prediction for combinations of process parameters that have not been trained on, and can automatically invert complex anisotropic rate fields, making it applicable to various surface processing problems such as etching and material growth.
[0006] Technical solution: Firstly, a method for constructing a meta-model of the evolution of micro / nano structure processing interfaces based on a physical information neural network constrained by Level-set equations, comprising the following steps:
[0007] Step 1: Based on the Level-set fundamental equations and the characteristic equations of the symbolic distance function, construct a set of partial differential equations describing the interface evolution; the symbolic distance function is a function of spatiotemporal coordinates and process parameter vectors, used to characterize the material interface position during the micro / nano structure fabrication process; this set of partial differential equations is equipped with corresponding initial conditions and boundary conditions;
[0008] Step 2: Construct a meta-model architecture that couples a main neural network and an auxiliary neural network. The main neural network takes spatiotemporal coordinates and process parameter vectors as input and outputs the corresponding symbolic distance function value. The auxiliary neural network takes local crystal plane normal vectors and process parameter vectors as input and outputs the full crystal plane normal rate. The local crystal plane normal vectors are obtained by automatically differentiating the symbolic distance function value output by the main neural network.
[0009] Step 3: Select a combination of process parameters within the preset process parameter range, process according to the selected combination of process parameters, collect interface geometric contour data after a specified processing time, and convert the interface geometric contour data into zero isosurface data of the signed distance function, which is used as the experimental observation true value in the training process.
[0010] Step 4: Substitute the outputs of the main neural network and the auxiliary neural network and their automatic differentiation results into the system of partial differential equations to calculate the physical constraint residuals; calculate the experimental data residuals based on the difference between the experimental observation true values and the output of the main neural network; construct a composite loss function using the physical constraint residuals and the experimental data residuals.
[0011] Step 5: By minimizing the composite loss function, the parameters of the main neural network and the auxiliary neural network are updated collaboratively, so that the network output approximates the experimental observation true value while satisfying the physical evolution law represented by the partial differential equation system. This yields a trained meta-model of the micro / nano structure processing interface evolution. This trained meta-model of the micro / nano structure processing interface evolution is used to predict the material interface evolution morphology during the micro / nano structure processing under given process parameters and spatiotemporal conditions.
[0012] Furthermore, based on the Level-set fundamental equations and the characteristic equations of the symbolic distance function, a set of partial differential equations describing the interface evolution is constructed, expressed as:
[0013]
[0014] in, The symbolic distance function, This represents the partial derivative of the signed distance function with respect to time. For the normal velocity of the entire crystal plane, For local crystal plane normal vectors, For process parameter vectors, The viscosity coefficient, For the Laplace operator, express The gradient.
[0015] Furthermore, the auxiliary neural network adopts a dual-branch coupled architecture, including a geometric feature branch and a trend feature branch;
[0016] The geometric feature branch takes the local crystal plane normal vector as input and outputs geometric features used to characterize the topological regularity of the rate map.
[0017] The trend feature branch takes the process parameter vector as input and outputs a scaling factor and a bias term;
[0018] Multiplying the scaling factor by the geometric feature and then adding it to the bias term yields the full-plane normal rate.
[0019] Furthermore, the full-plane normal rate output by the auxiliary neural network is substituted into the Level-set basic equation to calculate the Level-set equation residual. The Level-set equation residual is then automatically differentiated and fed back to the auxiliary neural network and the main neural network, respectively, so that the two constrain each other during training.
[0020] Furthermore, the output layer of the auxiliary neural network employs a linear activation function, enabling the value range of the full-plane normal rate to cover positive, negative, and zero real numbers. When the full-plane normal rate is greater than zero, the zero isosurface of the sign distance function advances along its positive normal direction to simulate the etching process of material removal. When the full-plane normal rate is less than zero, the zero isosurface of the sign distance function advances along its negative normal direction to simulate the deposition process of material growth. When the full-plane normal rate is equal to zero, the zero isosurface of the sign distance function remains unchanged to simulate the geometrically static state of the interface.
[0021] Furthermore, the physical constraint residuals include: Level-set equation residuals calculated from the Level-set fundamental equations, characteristic equation residuals calculated from the sign distance function characteristic equations, initial condition residuals calculated from the initial conditions, and boundary condition residuals calculated from the boundary conditions.
[0022] A composite loss function is constructed using the physical constraint residuals and the experimental data residuals, expressed as:
[0023]
[0024] In the formula, This represents the partial differential equation loss composed of the residuals of the Level-set equations and the residuals of the characteristic equations. Represents the initial condition residual. Represents the boundary condition residuals. Represents the residuals of experimental data. The weighting coefficients represent the losses in partial differential equations. The weighting coefficients represent the initial condition residuals. The weighting coefficients represent the boundary condition residuals. The weighting coefficient represents the residual of the experimental data.
[0025] Furthermore, the initial condition residual is obtained in the following manner:
[0026] The solution space is meshed, and the true symbolic distance function value at each mesh location is determined based on the geometry of the structure before processing and the corresponding material properties.
[0027] The initial spatiotemporal coordinates and process parameters are input into the main neural network, and the difference between the output of the main neural network and the true symbolic distance function value at the corresponding grid position is used as the initial condition residual.
[0028] Furthermore, the boundary condition residuals are obtained in the following manner:
[0029] Boundary sampling points are determined at the boundaries of the solution space region by random sampling or uniform sampling.
[0030] The spatiotemporal coordinates and process parameters at the boundary sampling points are input into the main neural network. The symbolic distance function within the mask coverage area does not change with processing time as a boundary constraint. The difference between the output of the main neural network at that point and the preset symbolic distance function value of the corresponding boundary sampling point is used as the boundary condition residual.
[0031] Secondly, this invention proposes a method for predicting the evolution of micro / nano structure fabrication interfaces, including:
[0032] A method for constructing a meta-model of the micro / nano structure processing interface evolution based on a physical information neural network constrained by Level-set equations is adopted to obtain a trained meta-model of the micro / nano structure processing interface evolution.
[0033] The spatiotemporal coordinates and process parameter vectors to be predicted are input into the main neural network of the micro / nano structure processing interface evolution meta-model, and the corresponding symbolic distance function value is output.
[0034] The local crystal plane normal vector obtained by automatic differentiation of the symbol distance function value is input into the auxiliary neural network to obtain the full crystal plane normal rate under the corresponding process parameters;
[0035] Substituting the full-crystal plane normal rate into the Level-set fundamental equation, the evolution of the sign distance function over time is iteratively solved, and the zero isosurface of the sign distance function during the iteration process is output as the evolution profile of the micro / nano structure processing interface.
[0036] Thirdly, this invention proposes a method for analyzing the normal rate map of the entire crystal plane, including:
[0037] A method for constructing a meta-model of the micro / nano structure processing interface evolution based on a physical information neural network constrained by Level-set equations is adopted to obtain a trained meta-model of the micro / nano structure processing interface evolution.
[0038] Set target process parameters and iterate through multiple preset crystal orientations. Input the local crystal orientation normal vectors corresponding to each crystal orientation into an auxiliary neural network to obtain the full crystal orientation normal rate under the target process parameters.
[0039] Using crystal orientation as the independent variable and the corresponding full-plane normal rate as the dependent variable, the full-plane normal rates corresponding to each crystal orientation are correlated to generate a full-plane normal rate spectrum under the target process parameters.
[0040] Based on the differences in the normal rates of the entire crystal planes between different crystal plane orientations in the full-crystal plane normal rate spectrum, the degree of anisotropy evolution of the micro / nano structure under the process parameters is evaluated.
[0041] Beneficial effects: Compared with the prior art, the present invention directly predicts the microstructure interface evolution from process parameters, constructs a meta-model that can make high-precision predictions of process parameter combinations that do not appear in the training set, and realizes anisotropic full-crystal rate inversion; it introduces an auxiliary neural network and adopts a unique dual-branch coupling architecture, which significantly improves the training efficiency and generalization of the model under small sample data; at the same time, it unifies the modeling of etching and deposition processes and achieves adaptive learning of the sign of normal rate. Attached Figure Description
[0042] Figure 1 This is a schematic diagram of a method for constructing a meta-model of the micro / nano structure processing interface based on a physical information neural network constrained by Level-set equations, as proposed in this invention.
[0043] Figure 2 This is a schematic diagram of the silicon substrate to be etched in Example 2;
[0044] Figure 3 A schematic diagram of an auxiliary neural network employing a dual-branch coupling architecture;
[0045] Figure 4 This is a diagram illustrating the training process of the meta-model.
[0046] Figure 5 This is a schematic diagram of the three-dimensional etching of gallium nitride material in Example 3. Detailed Implementation
[0047] This invention relates to efficient contour prediction and automatic inversion of full-crystal surface rate for various surface processing problems such as semiconductor etching and thin film deposition under different process conditions. Based on a physical information neural network, an end-to-end meta-model is established. By constraining the training of the neural network with a set of partial differential equations, initial conditions, boundary conditions, and small sample experimental data points, the invention achieves efficient and accurate prediction of the processing contour of materials under different process parameters and can invert the micro-etching rate.
[0048] To make the objectives, technical solutions, and advantages of this invention clearer, the following description, in conjunction with the accompanying drawings and embodiments, will further explain the method for constructing the meta-model of micro / nano structure processing interface based on a physical information neural network constrained by Level-set equations.
[0049] Example 1:
[0050] This invention proposes a method for constructing a meta-model of the micro / nanostructure fabrication interface evolution based on a physical information neural network constrained by Level-set equations. The constructed meta-model can achieve efficient contour prediction and automatic inversion of full-crystal rate for various surface processing problems, including etching and deposition, under different process conditions. The method of this invention achieves high-precision prediction of interface evolution by jointly constraining the training of the neural network with a set of partial differential equations, initial conditions, boundary conditions, and small sample experimental data points. Specifically, it includes the following steps:
[0051] Step 1: Construct a physical constraint framework based on residual control, including:
[0052] Define the residuals of the evolution equation:
[0053] To enhance numerical stability, the Level-set fundamental equations used to describe the evolution of anisotropic interfaces, after adding a viscosity term, are as follows:
[0054]
[0055] in, The symbolic distance function, The viscosity coefficient, Let be the partial derivative of the signed distance function with respect to time. Indicates the etching rate across the entire crystal plane. For the Laplace operator, express gradient, The Euclidean norm of the gradient is denoted as .
[0056] At the same time, the full-crystal etching rate Treat it as a parameterized function, and represent it as ,in, For local crystal plane normals, This is a vector of process parameters, including etching temperature. Etching solution concentration The influence of factors such as, are expressed as .
[0057] The residuals of the Level-set equations are:
[0058]
[0059] Define geometric residuals:
[0060] Addressing the issue of the signed distance function in traditional Level-set numerical solutions To address the issue of degradation over time, embodiments of the present invention directly construct characteristic equations. This is transformed into the residual of the characteristic equation:
[0061]
[0062] This constraint, acting as a regularization term, forces the signed distance function to maintain good geometric distance properties. In the 2D case, In 3D cases, , , and These are the signed distance functions. The partial derivatives with respect to spatial coordinates x, y, and z.
[0063] Therefore, the system of partial differential equations can be expressed as:
[0064]
[0065] The sign distance function of the solution of this system of partial differential equations It is about The function, the partial differential residual term of the physical constraint frame. Depend on and Appropriate weights are used to represent them together.
[0066] Step 2: To achieve unified modeling and rate inversion of interface evolution under different process parameters, a coupled architecture consisting of a neural network and an auxiliary neural network is constructed. Process parameters include, but are not limited to, etchant concentration and etching temperature.
[0067] Main Neural Network Used for modeling the symbolic distance function Simulate the interface evolution during the processing. Vectorize the process parameters. Additional features are used as input to predict continuous changes in process conditions, forming a meta-model.
[0068] Main Neural Network The input is: spacetime coordinates With process parameter vector .
[0069] Main Neural Network The output is: the symbolic distance function .
[0070] The network structure chosen is a simple fully connected neural network.
[0071] Due to the full-plane etching rate For a complex function of process parameters, an auxiliary neural network is introduced to implicitly learn and optimize this unknown function using data-driven learning. Modeling the etching rate of the entire crystal surface This auxiliary neural network, through coupling with the main neural network, operates based on the local crystal plane normal. With process parameter vector Output crystal plane rate values to achieve full-plane etching rate inversion. Specifically, the local crystal plane normal... By the signed distance function express:
[0072]
[0073] Simultaneously allow Output any real number to achieve unified processing of various surface finishing processes (etching, deposition, etc.). Specifically, if... Simulate the material removal process; if Simulate the material growth process; if The geometric static state of the simulated interface.
[0074] Auxiliary Neural Network The input is: local crystal plane normal. With process parameter vector .
[0075] Auxiliary Neural Network The output is: full-plane etching rate .
[0076] Based on the material's crystal structure and thermodynamic laws, the etching rate is decoupled into geometric anisotropy characteristics and physical evolution trend characteristics. (Auxiliary neural network) The architecture uses two parallel sub-fully connected neural networks: a geometric feature branch and a trend feature branch. The input to the geometric feature branch is... The output is a tensor. The topological shape of the characteristic reaction rate map; the input to the trend feature branch is a vector of process parameters. The output is a scaling factor. With bias , used to represent the effects of global scaling and baseline offset.
[0077] The final output definition is as follows:
[0078]
[0079] Step 3: Build a meta-model using small sample experimental data, including:
[0080] Experimental data acquisition under key process parameters S3-1:
[0081] Setting process parameter boundaries , A small sample sampling strategy is adopted, selecting only a small number of representative combinations within the process parameter boundaries. Perform etching or deposition experiments. For example, collect data for a specified etching time. The interface 3D point data is then converted into zero isosurface data of the sign distance function. This data serves as the real physical constraint for model training and has a dual function: on the one hand, it serves as a data constraint for predicting the processing profile; on the other hand, it serves as the observed true value for the inverse solution of the whole crystal etching rate.
[0082] S3-2 Custom Loss Function:
[0083] The loss function of this model consists of four residuals, including a partial differential equation loss composed of the residuals of the Level-set equations and the residuals of the characteristic equations. Initial condition loss Boundary condition loss and data constraint loss , Also known as experimental data residuals, the mean squared error is used as the basic form of the loss function. Since the magnitudes of different loss terms differ, appropriate weighting is required for each part. Therefore, the total loss is expressed as:
[0084]
[0085] in, The weighting coefficients represent the losses in partial differential equations. The weighting coefficients represent the initial condition residuals. The weighting coefficients represent the boundary condition residuals. The weighting coefficient represents the residual of the experimental data.
[0086] Main Neural Network Network parameters With auxiliary neural networks Network parameters Based on the total loss using the optimizer (Adam) The gradient changes are updated simultaneously, thus completing the process of training multiple related network modules through a unified loss function.
[0087] The specific losses are as follows:
[0088] Partial differential equation loss It is constructed based on the residuals of the Level-set equations and the residuals of the characteristic equations, and is expressed as:
[0089]
[0090] in, The number of sampling points within the microstructure region to be calculated. The partial differential residuals of the physical constraint framework are expressed as:
[0091]
[0092] in, These are the weighting coefficients for the residuals of the Level-set equations; These are the weighting coefficients of the residuals in the characteristic equation.
[0093] Initial condition loss , represented as:
[0094]
[0095] in, The initial number of sampling points is determined based on the initial interface complexity. The initial number of sampling points should be sufficient to adequately describe the geometric features of the initial interface. The initial constraint is reflected in the initialization of the signed distance function. The solution space domain is meshed, independent of process parameters; only the signed distance function value for each mesh needs to be determined based on the structural morphology and corresponding material properties. Therefore, the solution space domain can be used in... The signed distance function data constraint at time step 1 reflects the initial conditions of the PINN model. Therefore, the residuals of the initial conditions are expressed as:
[0096]
[0097] In the formula, the coordinates of the i-th initial condition sampling point in the spatial domain are... These represent the spatial sampling coordinates in the i-th x, y, and z directions, respectively. Let be the true symbolic distance function value corresponding to the i-th sampling point.
[0098] Boundary condition loss :
[0099]
[0100] in, The number of boundary point samples is obtained through random or uniform sampling at the boundary of the solution space region, fully describing the boundary geometry. Boundary condition constraints are embodied in the mask region, i.e., within the region covered by the mask, the signed distance function does not change with processing time t. Therefore, the boundary conditions of the PINN model can be represented by the spatiotemporal data constraints in the mask region. The boundary condition residuals are expressed as follows:
[0101]
[0102] In the formula, Indicates the etching time.
[0103] Data constraint loss , represented as:
[0104]
[0105] in, The number of data constraint sampling points is determined based on the complexity of the experimental profile, the temporal sampling density, and the model training accuracy requirements. This is based on a combination of various process parameters. Etching time The obtained geometric contour data is used as the signed distance function. Data constraints at the location. The data constraint residuals are represented as follows:
[0106]
[0107] It is the key to inversion, forcing the network to find the correct... To match the final outline.
[0108] S3-3 Collaborative training based on loss function:
[0109] Gradient descent is used to perform end-to-end co-training of the main neural network and the auxiliary neural network. Before training, the network parameters (weights and thresholds) of the main and auxiliary neural networks are initialized. The Adam optimizer is used for fast convergence training, followed by high-precision fine-tuning using the L-BFGS optimizer. The necessary hyperparameters and learning rate are then set. Maximum number of iterations Batch size Convergence accuracy of the loss value wait.
[0110] In each iteration, the main neural network The distribution of the output signed distance function is obtained, and the local normal vector is obtained through the signed distance function based on automatic differentiation, using an auxiliary neural network. The growth rate of the entire crystal plane is obtained. The loss value is calculated based on the residual module. If... If the training ends, the meta-model parameters are saved, indicating that the model has simultaneously mastered the interface evolution law and the ability to invert crystal surface rates; otherwise, the parameters of the main neural network and the auxiliary neural network are updated collaboratively.
[0111] Multidimensional predictions using the S3-4 gram model:
[0112] Using a trained meta-model, the interface evolution process and full-crystal rate corresponding to different process parameters within the solution range can be predicted, applicable to various processing methods. These include:
[0113] Interface morphology prediction across process parameters: Inputting previously unexperimented process parameters into a trained meta-model The main neural network directly outputs the corresponding signed distance function field, which can output the machining contour under the given process conditions within seconds, achieving high-precision prediction.
[0114] Generation of all-plane rate maps: Inputting omnidirectional normal phasors into the auxiliary neural network and process parameters The network directly outputs the corresponding full-plane velocity distribution data, realizing the inverse velocity calculation.
[0115] Unified processing of multiple processing methods: Due to the unified processing of multiple surface processing processes (etching, deposition, etc.) by the auxiliary neural network, this prediction process is not only applicable to anisotropic wet etching, but also to processing scenarios such as thin film deposition.
[0116] This invention relates to a method for solving interface changes in various surface processing problems such as semiconductor etching and thin film deposition under different process conditions. Based on a physical information neural network, an end-to-end meta-model is established to efficiently and accurately predict the processing profile of materials under different process parameters and to invert the micro-etching rate.
[0117] Example 2:
[0118] This invention addresses the 2D growth problem of silicon materials by using the construction method disclosed in Example 1 to obtain a meta-model of the micro / nano structure processing interface evolution, and performs profile prediction across process parameters.
[0119] like Figure 1 As shown, establishing a Level-set interface evolution meta-model for the growth problem generally includes the following steps:
[0120] The initial morphology of the material is divided into networks in the spatial domain to generate a training dataset for the initial time step of the symbolic distance function.
[0121] Training datasets based on boundary conditions generated from mask regions.
[0122] Determine the range of variation of process parameters (etching temperature and etching solution concentration), select a small number of representative combinations for growth experiments, record the morphology after a certain period of time, and generate a sparse signed distance function training dataset.
[0123] The Level-set fundamental equations and geometric property equations are embedded into the neural network as part of the loss function. The spatiotemporal domain and process parameters are used as inputs to the main neural network, and the symbolic distance function is used as the output. The local normal vector based on the symbolic distance function and process parameters are used as inputs to the auxiliary neural network, and the full crystal plane rate is used as the output. The trained neural network can regard the position where the symbolic distance function is equal to zero as the processing profile of the material, complete the prediction of interface changes under different process parameters, and output the change rate of different crystal planes.
[0124] The specific implementation steps are as follows:
[0125] Regarding the evolution of two-dimensional interfaces, a physical evolution model is established based on the Level-set fundamental equations and characteristic equations:
[0126] The basic equation for Level-set is: .
[0127] The characteristic equation of the level-set is: .
[0128] Therefore, the partial differential equations for solving the 2d interface growth problem are as follows:
[0129]
[0130] Determine the range of variation of the meta-model process parameters, conduct deposition experiments using representative parameter combinations, and record the processing profiles to generate a training dataset.
[0131] Determine the solution range for the interface evolution of silicon materials, with a length of [missing information]. Width is The predictable processing time range is Set the processing temperature With precursor concentration Take the interval. , , to obtain Under this combination Growth profile after time.
[0132]
[0133] Accurately measure and record the machining contours obtained from different combinations of process parameters. Take intervals. , , Divide the processing contour grid points:
[0134]
[0135]
[0136]
[0137] In the formula, ,
[0138] Grid points The signed distance function at the processing contour location is used as a reference value to form the training dataset. .
[0139] Determine the initial and boundary conditions for training the meta-model:
[0140] Determine the etching start time of the silicon substrate to be grown. The initial 2D geometry is used as the starting point for the evolution of the physical information neural network meta-model. This initial geometry is then transformed into a signed distance function. For any spatial point , The value represents the shortest distance from the point to the initial interface; positive values are found inside the interface, negative values are found outside, and zero values are found on the interface itself. The spatial point at the initial time and its corresponding Values, serving as data constraints for initial conditions.
[0141] The precise geometric position and shape of the mask layer on the silicon substrate surface to be etched are determined, represented as line segments in 2D. The mask region is considered as a fixed boundary. In the masked spatial domain, the symbolic distance function is constant and set as follows: This ensures that the area is always considered part of the material and remains unchanged. The spatial point located in the region and its fixed position Values serve as data constraints for boundary conditions.
[0142] like Figure 2 As shown in the figure, the white area is the solution region, which allows interface evolution; the black area is the mask position, and the white dashed outline represents the boundary position; the white solid line represents the material surface outline at the initial moment; the gray area is the silicon substrate material region.
[0143] Define a custom loss function for the physical information neural network meta-model and determine the structure of the neural network:
[0144] Using mean squared error, loss function It consists of four parts: partial differential equations, initial conditions, boundary conditions, and data constraints, which are consistent with the content of Example 1, and will not be described in detail here.
[0145] Main Neural Network With auxiliary neural networks All branches are fully connected neural networks. The input to the main neural network is... The output is the symbolic distance function. The hidden layer is set to Layers, each layer One neuron; such as Figure 3 As shown, the auxiliary neural network adopts a two-branch coupled architecture. The network consists of a geometric feature branch and a trend feature branch: the geometric feature branch only receives local crystal plane normals. As input, the output is used to represent the topological regularity of the rate map. This branch contains Each of the following fully connected hidden layers has a set of... One neuron; the trend feature branch only receives process parameter vectors. As input, the output is the scaling factor and bias term; this branch contains... Each of the following fully connected hidden layers has a set of... One neuron. The parameters output by the trend branch are combined with the features output by the geometric branch to finally output the rate of change of the entire crystal plane. The hidden layers of the network commonly employ... As an activation function to introduce nonlinearity, the output layer is typically linearly activated.
[0146] The neural network is trained based on the loss function and the model is saved:
[0147] The loss function measures the error between the neural network's output and a reference value. This error is then backpropagated, and the parameter weights in the network are continuously adjusted using gradient descent to reduce the error between the predicted and actual values. The specific steps are as follows: Figure 4 As shown:
[0148] Initialize the parameters (weights and thresholds) of the main neural network and auxiliary neural network, set the hyperparameters required for training, and the learning rate. Maximum number of iterations Batch size Convergence accuracy of the loss value wait.
[0149] A network optimizer is selected, and the distribution of the output symbolic distance function is obtained using the main neural network. Based on automatic differentiation, the local normal vector is obtained through the symbolic distance function, and the growth rate of the entire crystal surface is obtained using the auxiliary neural network.
[0150] Calculate the loss value ,like If the training ends, the meta-model parameters are saved; otherwise, the parameters of the main neural network and the auxiliary neural network are updated collaboratively.
[0151] By changing the process parameters, predictions can be made using a trained neural network meta-model.
[0152] Using the trained meta-model, different process parameters within the solution range can be predicted. The corresponding 2D interface growth evolution process and full-crystal growth rate. Input process parameters that have not been tested before. The main network can output the corresponding growth profile, and the auxiliary network can output the anisotropic growth rate map under this condition.
[0153] Example 3:
[0154] This invention addresses the three-dimensional etching problem of gallium nitride materials by performing full-plane rate inversion and morphology prediction based on a micro / nano structure processing interface evolution meta-model.
[0155] like Figure 1 As shown, the establishment of a Level-set interface evolution meta-model for the etching problem generally includes the following steps:
[0156] The initial morphology of the material is divided into networks in the spatial domain to generate a training dataset for the initial time step of the signed distance function.
[0157] Training datasets based on boundary conditions generated from mask regions.
[0158] Determine the range of variation of process parameters (etching temperature and etching solution concentration), select a small number of representative combinations for growth experiments, record the morphology after a certain period of time, and generate a sparse signed distance function training dataset.
[0159] The Level-set fundamental equations and characteristic equations are embedded into a neural network as part of the loss function. The spatiotemporal domain and process parameters are used as inputs to the main neural network, and the symbolic distance function is used as the output. The local normal vector based on the symbolic distance function and process parameters are used as inputs to the auxiliary neural network, and the full crystal plane rate is used as the output. The trained neural network can regard the position where the symbolic distance function is equal to zero as the processing profile of the material, complete the prediction of interface changes under different process parameters, and output the change rate of different crystal planes.
[0160] The specific implementation steps are as follows:
[0161] Regarding the evolution of 3D interface etching, a physical evolution model is established based on the Level-set fundamental equations and characteristic equations:
[0162] The basic equation for Level-set is: .
[0163] The characteristic equation of the level-set is: .
[0164] Therefore, the solution to the partial differential equations for the three-dimensional interface etching problem is as follows:
[0165]
[0166] in, The sign distance function in three-dimensional space. The etching rate is the normal to the entire crystal plane (for the etching process, ).
[0167] Determine the range of variation of the meta-model process parameters, conduct wet etching experiments using representative parameter combinations, and record the final processing contour to generate a training dataset.
[0168] Determine the solution range for the 3D interface evolution of GaN materials, with a length of [missing information]. Width is The height is The predictable processing time range is Set the processing temperature. With precursor concentration Take the interval. , , to obtain Under this combination of time The final etched outline.
[0169] Accurately measure and record the final 3D machining profile obtained from different combinations of process parameters. (Take intervals.) , , Divide the processing contour grid points:
[0170]
[0171]
[0172] ,
[0173] In the formula, , , .
[0174] The signed distance function of the grid points at the final processed contour position is used as a reference value to form the training dataset. This data is the core ground truth observation for driving rate inversion.
[0175] Determine the initial and boundary conditions for training the meta-model:
[0176] Determine the starting time of the GaN substrate to be etched. The initial three-dimensional geometric shape is used as the starting point for the evolution of the PINN meta-model. This initial shape is then transformed into a signed distance function. For any spatial point , The value represents the shortest distance from that point to the initial interface; positive values are used inside the interface, negative values are used outside, and zero values are used on the interface itself. The spatial point at the initial time and its corresponding Values, serving as data constraints for initial conditions.
[0177] The precise geometric position and shape of the mask layer on the GaN substrate surface are determined, treating the mask region as a fixed boundary. In the masked spatial domain, the symbolic distance function is constant and set as follows: This ensures that the area remains visible and that the material's interior remains unchanged. The spatial point located in the mask area and its fixed position Values serve as data constraints for boundary conditions.
[0178] like Figure 5 As shown, the solution domain It is a three-dimensional cuboid space. The mask area is a specific covered area (black) on the top surface of the hexagonal frustum, and the etchable area is the exposed area under the mask opening (the side of the hexagonal frustum). The hexagonal frustum is the etched GaN material.
[0179] Define a custom loss function for the physical information neural network meta-model and determine the structure of the neural network:
[0180] Using mean squared error, loss function It consists of four parts: a system of partial differential equations, initial conditions, boundary conditions, and data constraints. The content is consistent with Example 1, and will not be elaborated further here.
[0181] Both the main neural network and the auxiliary neural network have fully connected branch structures. The input to the main neural network is... The output is the symbolic distance function. The hidden layer is set to Layers, each layer One neuron; such as Figure 3 As shown, the auxiliary neural network adopts a two-branch coupled architecture. The network consists of a geometric feature branch and a trend feature branch: the geometric feature branch only receives local crystal plane normals. As input, the output is used to represent the topological regularity of the rate map. This branch contains Each of the following fully connected hidden layers has a set of... One neuron; the trend feature branch only receives process parameter vectors. As input, the output is the scaling factor and bias term; this branch contains... Each of the following fully connected hidden layers has a set of... One neuron. The parameters output by the trend branch are combined with the features output by the geometric branch to ultimately output the rate of change of the entire crystal plane. The hidden layers of the network commonly employ... As an activation function to introduce nonlinearity, the output layer is typically linearly activated.
[0182] The neural network is trained based on the loss function and the model is saved:
[0183] The loss function measures the error between the neural network's output and a reference value. This error is then backpropagated, and the parameter weights in the network are continuously adjusted using gradient descent to reduce the error between the predicted and actual values. The specific steps are as follows: Figure 4 As shown:
[0184] Initialize the parameters (weights and thresholds) of the main neural network and auxiliary neural network, set the hyperparameters required for training, and the learning rate. Maximum number of iterations Batch size Convergence accuracy of the loss value wait.
[0185] A network optimizer is selected, and the distribution of the output symbolic distance function is obtained using the main neural network. Based on automatic differentiation, the local normal vector is obtained through the symbolic distance function, and the growth rate of the entire crystal surface is obtained using the auxiliary neural network.
[0186] Calculate the loss value ,like If the training ends, the meta-model parameters are saved; otherwise, the parameters of the main neural network and the auxiliary neural network are updated collaboratively, and the process proceeds to the second loop.
[0187] By changing the process parameters, predictions can be made using a trained neural network meta-model.
[0188] Using a trained meta-model, arbitrary process parameters Input the omnidirectional unit normal vector into the auxiliary neural network The network can then output a complete three-dimensional full-crystal etching rate map under these conditions. Simultaneously, the main neural network can predict etching profiles under different process conditions.
Claims
1. A method for constructing a meta-model of the evolution of micro / nanostructure fabrication interfaces based on a physical information neural network constrained by Level-set equations, characterized in that, Includes the following steps: Step 1: Based on the Level-set fundamental equations and the characteristic equation of the symbolic distance function, construct a set of partial differential equations describing the interface evolution; the symbolic distance function is a function of spatiotemporal coordinates and process parameter vectors, used to characterize the material interface position during the micro / nano structure fabrication process; this set of partial differential equations is equipped with corresponding initial and boundary conditions. Step 2: Construct a meta-model architecture coupling a main neural network and an auxiliary neural network. The main neural network takes spatiotemporal coordinates and process parameter vectors as input and outputs the corresponding symbolic distance function value; the auxiliary neural network takes local crystal plane normal vectors and process parameter vectors as input and outputs the full crystal plane normal rate; wherein, the local crystal plane normal vector is obtained by automatically differentiating the symbolic distance function value output by the main neural network. Step 3: Select a combination of process parameters within the preset process parameter range, process according to the selected combination of process parameters, collect interface geometric contour data after a specified processing time, and convert the interface geometric contour data into zero isosurface data of the signed distance function, which is used as the experimental observation true value in the training process. Step 4: Substitute the outputs of the main neural network and the auxiliary neural network and their automatic differentiation results into the system of partial differential equations to calculate the physical constraint residuals; calculate the experimental data residuals based on the difference between the experimental observation true values and the output of the main neural network; construct a composite loss function using the physical constraint residuals and the experimental data residuals. Step 5: By minimizing the composite loss function, the parameters of the main neural network and the auxiliary neural network are updated collaboratively, so that the network output approximates the experimental observation true value while satisfying the physical evolution law represented by the partial differential equation system. This yields a trained meta-model of the micro / nano structure processing interface evolution. This trained meta-model of the micro / nano structure processing interface evolution is used to predict the material interface evolution morphology during the micro / nano structure processing under given process parameters and spatiotemporal conditions.
2. The method for constructing a meta-model of the micro / nano structure fabrication interface evolution based on a physical information neural network constrained by Level-set equations, as described in claim 1, is characterized in that... The partial differential equations describing the interface evolution, based on the Level-set fundamental equations and the characteristic equations of the symbolic distance function, are constructed as follows: ; in, The symbolic distance function, This represents the partial derivative of the signed distance function with respect to time. For the normal velocity of the entire crystal plane, For local crystal plane normal vectors, For process parameter vectors, The viscosity coefficient, For the Laplace operator, express gradient, The Euclidean norm of the gradient is represented.
3. The method for constructing a meta-model of the micro / nano structure fabrication interface evolution based on a physical information neural network constrained by Level-set equations, as described in claim 1, is characterized in that... The auxiliary neural network adopts a dual-branch coupled architecture, including a geometric feature branch and a trend feature branch; The geometric feature branch takes the local crystal plane normal vector as input and outputs geometric features used to characterize the topological regularity of the rate map. The trend feature branch takes the process parameter vector as input and outputs a scaling factor and a bias term; Multiplying the scaling factor by the geometric feature and then adding it to the bias term yields the full-plane normal rate.
4. The method for constructing a meta-model of the micro / nanostructure fabrication interface evolution based on a physical information neural network constrained by Level-set equations, as described in claim 3, is characterized in that... The full-plane normal rate output by the auxiliary neural network is substituted into the Level-set basic equation to calculate the Level-set equation residual. The Level-set equation residual is then automatically differentiated and fed back to the auxiliary neural network and the main neural network, respectively, so that the two constrain each other during training.
5. The method for constructing a meta-model of the micro / nanostructure fabrication interface evolution based on a physical information neural network constrained by Level-set equations, as described in claim 3, is characterized in that... The output layer of the auxiliary neural network uses a linear activation function to make the range of values for the full-plane normal rate cover positive, negative, and zero real numbers. When the full-plane normal rate is greater than zero, the zero isosurface of the sign distance function advances along its positive normal direction to simulate the etching process of material removal. When the full-plane normal rate is less than zero, the zero isosurface of the sign distance function advances along its negative normal direction to simulate the deposition process of material growth. When the full-plane normal rate is equal to zero, the zero isosurface of the sign distance function remains unchanged to simulate the geometric static state of the interface.
6. The method for constructing a meta-model of the evolution of micro / nano structure processing interfaces based on a physical information neural network constrained by Level-set equations, as described in claim 1, is characterized in that... The physical constraint residuals include: Level-set equation residuals calculated from the Level-set fundamental equations, characteristic equation residuals calculated from the characteristic equation of the sign distance function, initial condition residuals calculated from the initial conditions, and boundary condition residuals calculated from the boundary conditions. A composite loss function is constructed using the physical constraint residuals and the experimental data residuals, expressed as: ; In the formula, This represents the partial differential equation loss composed of the residuals of the Level-set equations and the residuals of the characteristic equations. Represents the initial condition residual. Represents the boundary condition residuals. Represents the residuals of experimental data. The weighting coefficients represent the losses in partial differential equations. The weighting coefficients represent the initial condition residuals. The weighting coefficients represent the boundary condition residuals. The weighting coefficient represents the residual of the experimental data.
7. The method for constructing a meta-model of the micro / nanostructure fabrication interface evolution based on a physical information neural network constrained by Level-set equations, as described in claim 6, is characterized in that... The initial condition residuals are obtained in the following manner: The solution space is meshed, and the true symbolic distance function value at each mesh location is determined based on the geometry of the structure before processing and the corresponding material properties. The initial spatiotemporal coordinates and process parameters are input into the main neural network, and the difference between the output of the main neural network and the true symbolic distance function value at the corresponding grid position is used as the initial condition residual.
8. The method for constructing a meta-model of the micro / nanostructure fabrication interface evolution based on a physical information neural network constrained by Level-set equations, as described in claim 6, is characterized in that... The boundary condition residuals are obtained in the following way: Boundary sampling points are determined at the boundaries of the solution space region by random sampling or uniform sampling. The spatiotemporal coordinates and process parameters at the boundary sampling points are input into the main neural network. The symbolic distance function within the mask coverage area does not change with processing time as a boundary constraint. The difference between the output of the main neural network at that point and the preset symbolic distance function value of the corresponding boundary sampling point is used as the boundary condition residual.
9. A method for predicting the evolution of fabrication interfaces in micro / nano structures, characterized in that, include: The method for constructing a meta-model of micro / nano structure processing interface evolution based on a physical information neural network constrained by Level-set equations, as described in any one of claims 1 to 8, yields a trained meta-model of micro / nano structure processing interface evolution. The spatiotemporal coordinates and process parameter vectors to be predicted are input into the main neural network of the micro / nano structure processing interface evolution meta-model, and the corresponding symbolic distance function value is output. The local crystal plane normal vector obtained by automatic differentiation of the symbol distance function value is input into the auxiliary neural network to obtain the full crystal plane normal rate under the corresponding process parameters; Substituting the full-crystal plane normal rate into the Level-set fundamental equation, the evolution of the sign distance function over time is iteratively solved, and the zero isosurface of the sign distance function during the iteration process is output as the evolution profile of the micro / nano structure processing interface.
10. A method for analyzing the normal rate map of a whole crystal plane, characterized in that, include: The method for constructing a meta-model of micro / nano structure processing interface evolution based on a physical information neural network constrained by Level-set equations, as described in any one of claims 1 to 8, yields a trained meta-model of micro / nano structure processing interface evolution. Set target process parameters and iterate through multiple preset crystal orientations. Input the local crystal orientation normal vectors corresponding to each crystal orientation into an auxiliary neural network to obtain the full crystal orientation normal rate under the target process parameters. Using crystal orientation as the independent variable and the corresponding full-plane normal rate as the dependent variable, the full-plane normal rates corresponding to each crystal orientation are correlated to generate a full-plane normal rate spectrum under the target process parameters. Based on the differences in the normal rates of the entire crystal planes between different crystal plane orientations in the full-crystal plane normal rate spectrum, the degree of anisotropy evolution of the micro / nano structure under the process parameters is evaluated.
