A linear equation set self-adaptive optimization solving method and system for geotechnical medium mechanics simulation

Abstract

This invention discloses an adaptive optimization solution method and system for linear equation systems in geotechnical mechanics simulation, belonging to the interdisciplinary fields of geotechnical engineering numerical simulation, sparse linear equation system solution, and intelligent computing. The algorithm extracts the coefficient matrix and right-hand side vector of the actual solution required for the current calculation step from the geotechnical mechanics simulation program. It then classifies and saves training samples based on residual level, error threshold, number of linear iterations, nonlinear or contact state iteration information, time step or load increment adjustment status, and fixed calculation step interval. This invention avoids efficiency fluctuations caused by selecting solvers solely based on empirical rules or fixed methods, improving the solution efficiency, stability, and adaptability in complex geotechnical simulation processes.

Description

Technical Field

[0001] This invention relates to the field of data processing technology for simulation of soil and rock media mechanics, and more specifically, to an adaptive optimization solution method and system for linear equation systems used in soil and rock media mechanics simulation. Background Technology

[0002] In the finite element method, discrete element method, numerical manifold method, discontinuous deformation analysis method and their coupled simulation processes, it is usually necessary to repeatedly solve large-scale sparse linear equation systems in each calculation step, load step or nonlinear iteration step. As the computational scale increases, the contact relationships become more complex and the material nonlinearity increases, the solution efficiency and convergence stability of the linear equation systems gradually become the key factors restricting the overall efficiency of geotechnical mechanics simulation. Summary of the Invention

[0003] To address the technical problems mentioned above, this invention provides an adaptive optimization solution method and system for linear equation systems used in geotechnical mechanics simulation.

[0004] This invention provides an adaptive optimization solution method for linear equation systems used in geotechnical mechanics simulation, comprising the following steps: S1. From the geotechnical mechanics simulation data processing, obtain the linear equation system data that actually needs to be solved at the current time step. The linear equation system data includes the coefficient matrix. ,vector and the right-hand term vector ,in, The coefficient matrix is ​​assembled at the current time step. Let be the vector to be solved. It is the right-hand term vector formed by external loads, inertia terms, damping terms, constraint terms, and / or contact action terms; S2. Based on the residual level, error threshold, number of linear iterations, nonlinear or contact state iteration information, time step or load increment adjustment state, and fixed calculation step interval during the simulation data processing, classify and save the linear equation system data to form a training matrix sample library. S3. Based on the sparse structure, diagonal dominance, symmetry, ill-conditionedness, substructure coupling, right-hand side state, and computational state of linear equation systems, a matrix characteristic index system is constructed to characterize the difficulty of solving linear equation systems, and the coefficient matrix is ​​analyzed. and the right-hand term vector Feature extraction is performed to obtain the matrix eigenvectors; S4. Construct a set of candidate solvers and a set of candidate preprocessors. For each sample in the training matrix sample library, iterate through the combinations of candidate solvers and candidate preprocessors, and record the convergence state, solution time, number of iterations and relative residuals of each combination under the unified residual accuracy requirements and maximum solution time constraints. S5. Based on the solution results recorded in S4, determine the optimal solution strategy label and the suboptimal solution strategy label from the candidate combinations that meet the accuracy requirements. S6. Using the matrix eigenvectors obtained in S3 as input and the optimal solution strategy labels obtained in S5 as output, train the solution strategy selection model. S7. Embed the trained solution strategy selection model into the simulation data processing of soil and rock mechanics. In the process of simulation data processing, perform online feature extraction on the linear equation system of the new calculation step, and predict the corresponding solver-preprocessor combination through the solution strategy selection model. S8. Call the solver-preprocessor combination predicted in S7 to complete the solution of the current linear equation system and return the solution results to the simulation data processing flow. S9. Judge the convergence and solution efficiency of the solution results of the current linear equation system; when the predicted solver-preprocessor combination does not meet the preset residual accuracy requirements or exceeds the maximum solution time, call the suboptimal solution strategy or the preset baseline solver to complete the solution, and feed back the corresponding linear equation system samples and actual solution results to the training matrix sample library for subsequent model updates.

[0005] Among them, matrix eigenvectors Represented as:

[0006] In the formula, Represents the eigenvectors of a matrix; Indicates the order of a matrix; Indicates the number of active blocks; Representation matrix The number of non-zero elements in the middle; Represents matrix density; Represents the sparsity of a matrix; and These represent the mean, maximum, minimum, and standard deviation of the number of non-zero elements in a row, respectively. and These represent the mean, maximum, and standard deviation of the number of non-zero elements in the column, respectively. and These represent the upper bandwidth, lower bandwidth, and half bandwidth, respectively. , and These represent the mean of the absolute values ​​of the diagonal elements, the maximum absolute value of the diagonal elements, and the minimum absolute value of the non-zero diagonal elements, respectively. and These represent the zero diagonal scale and the near-zero diagonal scale, respectively. and These represent the proportion of diagonally dominant rows and the proportion of strictly diagonally dominant rows, respectively. and These represent the mean and maximum values ​​of the off-diagonal dominance ratio, respectively. This indicates a significant proportion of non-diagonal dominance; Represents matrix symmetry error; Represents an asymmetric, non-zero element proportion; , and Let them represent the matrix 1-norm, infinity norm, and... Norm; This represents the condition number estimate; express Block density; Indicates the number of non-zero 6×6 sub-blocks; Indicates block bandwidth; Indicates the proportion of non-diagonal blocks; Indicates the block coupling strength; Represents the sum of the diagonal block norms; Represents the sum of norms of off-diagonal blocks; , and They represent the right-hand term vectors respectively. The 1-norm, 2-norm, and infinity norm; and These represent the mean and standard deviation of the absolute values ​​of the terms on the right-hand side vector, respectively. This indicates that the right-hand side is a non-zero proportion; Indicates the initial residual; This represents the residual from the previous time step; Indicates the number of iterations in the previous time step; Indicates the current time step; A flag variable indicating whether time step adjustment is triggered.

[0007] The present invention also provides an adaptive optimization solution system for linear equations in geotechnical mechanics simulation, comprising a matrix data acquisition module, a sample classification module, a matrix feature extraction module, a candidate combination testing module, a label generation module, a model training module, an online prediction module, a solver scheduling module, and a sample backflow and model update module. The system is used to execute the above-described methods.

[0008] The beneficial effects of this invention are as follows: This invention addresses the linear equation systems actually formed during geotechnical mechanics simulations, constructing a multi-dimensional matrix characteristic index system that considers matrix structure, numerical characteristics, and computational state characteristics. Compared to traditional methods that rely solely on model size, number of blocks, single empirical rules, or fixed solver configurations, this invention comprehensively characterizes the difficulty of solving linear equation systems from multiple perspectives, including matrix size, non-zero element distribution, sparse structure, bandwidth, diagonal dominance, symmetry, ill-conditionedness, substructure coupling relationships, right-hand side state, and the current computational step's solution state. This results in a more refined, accurate, and interpretable basis for solution strategy selection. Particularly in the preferred DDA embodiment, this invention can also combine block structure and contact coupling evolution characteristics to more effectively characterize changes in matrix properties under complex contact conditions.

[0009] This invention transforms the problem of solving linear equations in geotechnical mechanics simulations from traditional manual judgment or fixed rule-based methods into a mapping problem between matrix characteristics and the performance of candidate solver-preprocessor combinations. By conducting systematic testing on multiple candidate solver-preprocessor combinations under unified residual accuracy requirements and time constraints, and generating optimal and suboptimal solution strategy labels based on actual convergence state, relative residuals, iteration count, and solution time, the trained model can directly learn towards the engineering goal of "reducing solution time while maintaining accuracy." Therefore, this invention not only improves the automation level of solver selection but also ensures that the model optimization objective aligns with the goal of improving simulation computational efficiency.

[0010] This invention proposes an online deployable adaptive optimization solution mechanism. During simulation, it automatically performs feature extraction, solution strategy prediction, and solver scheduling for each new computational step of the linear equation system. Compared to using a single solver or preprocessor for all cases, this invention dynamically switches to a solution strategy more suitable for the current equation system based on real-time changes in matrix features. This effectively reduces problems such as slow convergence, excessive iterations, or preprocessor mismatch on certain matrices, improving the overall solution efficiency and stability in complex nonlinear processes, contact evolution processes, and large-scale simulations.

[0011] This invention further retains the suboptimal solution strategy based on model prediction and sets up a fallback solution mechanism. When the predicted solver-preprocessor combination fails to meet the preset accuracy requirements, exceeds the preset maximum solution time, or fails to converge on the current linear equation system, it can automatically switch to the suboptimal combination or switch to the baseline solver preset by the simulation program to continue solving. This avoids the interruption of the overall calculation due to a single prediction error, significantly enhancing the stability, robustness, and fault tolerance of this invention in engineering applications, and is especially suitable for simulation scenarios where matrix properties change rapidly and the solution difficulty fluctuates greatly.

[0012] This invention offers good engineering embeddability and ease of implementation. The method does not alter the basic mechanical solution framework, governing equation form, or main computational flow of existing geotechnical mechanics simulation programs. Instead, it adds feature extraction, strategy prediction, and combined scheduling modules outside the original linear equation solution stage, thereby achieving optimized solution efficiency with minimal program modifications. Therefore, this invention is easily embedded into finite element method, discrete element method, numerical manifold method, discontinuous deformation analysis method, and their coupled simulation programs, exhibiting good versatility, scalability, and widespread application value.

[0013] This invention establishes a mechanism for continuous accumulation of training samples and iterative model updates. For samples that fail to converge, newly added samples, typical difficult matrices, or complex contact stage matrices encountered during the calculation process, these samples can be fed back into the training sample library. After supplementary testing and retraining, the solution strategy and model selection are continuously optimized. Therefore, this invention not only adapts to the differences between various types of geotechnical mechanics simulation problems, but also gradually improves the model's generalization ability and prediction accuracy as the sample size increases and the number of simulation scenarios grows, enabling this adaptive optimization solution algorithm to possess the ability for continuous evolution and long-term optimization. Attached Figure Description

[0014] Figure 1 This is a schematic diagram of the method flow of the present invention. Detailed Implementation

[0015] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0016] See Figure 1 This invention discloses an adaptive optimization solution method for linear equation systems used in geotechnical mechanics simulation.

[0017] In this embodiment, after completing stiffness assembly, constraint processing, contact processing, or nonlinear correction in each calculation step, the geotechnical mechanics simulation program forms the following set of linear equations: ; In the formula, For the first The coefficient matrix formed at each time step For the first A vector of unknowns to be solved at each time step; For the first The right-hand term vector at each time step; Number the time step.

[0018] It can be formed by the block material stiffness, boundary constraints, contact constraints, mass term, and damping term. It can be formed by external load, inertial term, damping term and state variables of the previous time step.

[0019] When the geotechnical mechanics simulation program performs discontinuous deformation analysis In programming, linear equation systems are typically formed after contact search, opening and closing iterations, and matrix assembly at each time step. Two-dimensional... In this context, each active block typically corresponds to 6 degrees of freedom. If the number of active blocks participating in the solution at the current time step is... Then the coefficient matrix order It can be represented as: ; In the formula, For the first Time step coefficient matrix The order of; This represents the number of active blocks participating in the solution at the current time step. In a preferred embodiment, if some blocks are in a dead, fixed, or non-participating state in the current solution, only the active blocks that actually participate in the solution are consecutively numbered to avoid empty rows, empty columns, or invalid degrees of freedom in the derived matrix.

[0020] In this embodiment, the linear equation samples generated during the simulation of soil and rock mechanics are classified and exported.

[0021] This indicates the residual level after solving the linear equation system in the current computational step. This indicates the preset error threshold. This indicates the number of iterations required to solve the system of linear equations.

[0022] Sample classification includes: When the residual Not greater than the error threshold When the time is right, the corresponding matrix is ​​denoted as the normally convergent matrix; when Greater than and not greater than When this happens, the corresponding matrix is ​​denoted as the slightly difficult matrix; when Greater than and not greater than When the matrix is ​​of medium difficulty, the corresponding matrix is ​​denoted as the matrix of medium difficulty. when Greater than When this happens, the corresponding matrix is ​​denoted as the difficult matrix; When the number of linear iterations When the preset high iteration threshold is reached, the corresponding matrix is ​​recorded as the high iteration matrix; When the simulation program triggers a recalculation by reducing the time step, the corresponding matrix is ​​recorded as the time step adjustment matrix; When the time step number meets the preset fixed interval condition, the corresponding matrix is ​​recorded as the fixed interval matrix.

[0023] This classification strategy allows the training sample library to cover various matrix states in simulation calculations, ranging from normal convergence to difficult solutions. In a preferred embodiment, the incremental adjustment matrix may correspond to the case where the program triggers a recalculation of the time step reduction.

[0024] Specifically, for the coefficient matrix and the right-hand term vector Feature extraction; time step subscripts are omitted below for simplicity. The current time step matrix is ​​denoted as Let the vector of the right-hand side be denoted as .

[0025] The matrix eigenvectors include at least one of the following: matrix size features, sparse structure features, row non-zero element distribution features, column non-zero element distribution features, bandwidth features, diagonal element features, diagonal dominance features, off-diagonal coupling features, symmetry features, ill-conditioned features, norm features, substructure features, substructure coupling features, right-hand term vector features, initial residual features, solution history features, and computational step state features.

[0026] Among them, the matrix size characteristics include at least one of the following: matrix order, number of degrees of freedom, number of active substructures, and number of non-zero elements; Sparse structure features include at least one of matrix density and sparsity; The distribution characteristics of row non-zero elements include at least one of the following: average number of row non-zero elements, maximum number of row non-zero elements, minimum number of row non-zero elements, and standard deviation of the number of row non-zero elements. The distribution characteristics of column non-zero elements include at least one of the following: average number of column non-zero elements, maximum number of column non-zero elements, and standard deviation of the number of column non-zero elements; Bandwidth characteristics include at least one of upper bandwidth, lower bandwidth, and half bandwidth; Diagonal element characteristics include at least one of the following: mean absolute value of diagonal elements, maximum absolute value of diagonal elements, minimum absolute value of non-zero diagonal elements, zero diagonal proportion, and near-zero diagonal proportion. Diagonal dominance features include at least one of the following: diagonal dominance row proportion, strict diagonal dominance row proportion, and weak diagonal dominance row proportion. Off-diagonal coupling characteristics include at least one of the following: mean off-diagonal dominance ratio, maximum off-diagonal dominance ratio, and severe off-diagonal dominance ratio. Symmetry features include at least one of matrix symmetry error, asymmetric non-zero element ratio, and upper and lower triangular non-zero element difference ratio; Illness characteristics include at least one of condition number estimation, diagonal element span, row and span, and column and span; Norm features include at least one of matrix norm ratio and near-singularity criterion; Substructure characteristics include at least one of the following: number of active substructures, number of non-zero sub-blocks, sub-block density, proportion of off-diagonal sub-blocks, sub-block bandwidth, and non-zero distribution statistics of sub-blocks; Substructure coupling characteristics include at least one of the following: sum of diagonal subblock norms, sum of off-diagonal subblock norms, substructure coupling strength, maximum off-diagonal subblock norm, average off-diagonal subblock norm, and number of coupled substructure pairs; The features of the right-hand term vector include at least one of the following: right-hand term 2 norm, right-hand term infinite norm, right-hand term 1 norm, right-hand term mean, right-hand term standard deviation, right-hand term maximum absolute value, right-hand term non-zero proportion, and right-hand term distribution skewness. The computational step state characteristics include at least one of the following: computational step number, current time step or load increment, whether incremental adjustment is triggered, number of iterations for nonlinear or contact states, number of linear iterations, and residual from the previous computational step.

[0027] set up ,in, Representation matrix No. Line number Column elements, express OK The set of real matrix columns and These represent the row number and column number of the matrix, respectively.

[0028] Then, the following features are extracted: Matrix density is expressed as , For matrix The number of non-zero elements, For matrix The order of; The sparsity of a matrix is ​​expressed as ; matrix No. The number of non-zero elements in a row is represented as , # indicates the number of elements in the set, Column numbering. By The average number of non-zero elements in a row, the maximum number of non-zero elements in a row, and the standard deviation of non-zero elements in a row can be further calculated to describe the uniformity of the distribution of non-zero elements in the current linear equation system matrix.

[0029] No. The absolute value of the diagonal elements of a row is expressed as ; For matrix A, the first... Line number The diagonal elements of the column; No. The sum of the absolute values ​​of the off-diagonal elements of a row is expressed as ; The diagonal dominance ratio is expressed as ; This is an indicator function that takes the value 1 if the condition within the parentheses is true, and 0 otherwise. No. The off-diagonal dominance ratio is expressed as ; To prevent extremely small positive numbers with a denominator of zero.

[0030] and It is used to describe the degree of diagonal dominance of a matrix and serves as an indicator of the convergence of iterative methods such as SOR, Jacobi, and Block Jacobi.

[0031] Matrix symmetric error is expressed as ; Let A be the transpose of matrix A; for Norm. The closer a matrix is ​​to 0, the more symmetric it is.

[0032] When the simulation program is During programming, the coefficient matrix also exhibits obvious block structure and contact coupling characteristics. For two-dimensional... Each active block typically corresponds to 6 degrees of freedom, therefore matrix A can be further divided into 6×6 submatrices and extracted. Block structure features and block coupling features: DDA block density is expressed as ; The number of non-zero 6×6 sub-blocks in A; This represents the number of active blocks.

[0033] Block coupling strength is expressed as ; For the first The block and the first Each block corresponds to a 6×6 submatrix. Used to describe the strength of contact coupling between blocks relative to the diagonal contribution of the blocks themselves.

[0034] The above features are used to characterize The sub-block connections generated by contact, constraints, and dynamic coupling between blocks help identify the changes in the matrix block structure at different time steps and the corresponding solution difficulty.

[0035] Right-hand term vector The 2-norm is represented as ; The right-hand term vector The i-th component; Right-hand term vector The infinite norm is expressed as ; The eigenvectors of a matrix are represented as ; It is a matrix eigenvector formed by concatenating features from multiple types of matrices; This represents the average number of non-zero elements in a row. This indicates the maximum number of non-zero elements in a row; the ellipsis indicates that extended features such as bandwidth, zero diagonal proportion, condition number estimate, mean and standard deviation of right-hand side terms may also be included. When applied to... During the program, the scene-specific features may further include block bandwidth, zero diagonal ratio, condition number estimation, contact coupling strength, and statistical features of right-hand terms, etc.

[0036] In this embodiment, a candidate solver set M and a candidate preprocessor set P are constructed. Set M includes at least one of GMRES, BiCGSTAB, FGMRES, PCG, MINRES, SOR, and Block Jacobi; set P includes at least one of None, Jacobi, Block Jacobi, IC, ILU, AMG, and SSOR.

[0037] For the The training sample, using the th training sample The solver and the first The solution is obtained by combining several preprocessors, and the solution vector is recorded. Solution time and relative residuals Among them, relative residual Represented as: ; In the formula, For the first The coefficient matrix of each training sample; For the first The right-hand term vector of each training sample; This is the solution vector obtained from this combination; It represents the 2-norm.

[0038] No. Ranking results of candidate effective combinations for each training sample Represented as: ; In the formula, Indicates according to solution time Sort by size from smallest to largest; The target residual threshold used for comparison; The value is set to 1 to indicate successful convergence, and 0 otherwise, it is set to 0.

[0039] The combination ranked first is denoted as the optimal solution strategy label, and the combination ranked second is denoted as the suboptimal solution strategy label, thus transforming the performance differences of different matrix features into training labels that can be used for supervised learning.

[0040] The training matrix sample library is represented as... ,in For the training dataset; For the first The matrix eigenvectors of each sample; For the first The optimal solution strategy label for each sample; This represents the number of training samples.

[0041] Model selection strategy: Use random forest, gradient boosting tree, support vector machine, neural network or other classification models to train and solve the problem. .

[0042] Solution strategy labels predicted by the model Represented as: ; In the formula, For parameters Solution strategy selection model; The set of parameters obtained during model training; This is the eigenvector of the current matrix.

[0043] During online deployment, the simulation program forms a matrix in the new computation step. and the right-hand term vector Then, extract the corresponding feature vectors. And through the model Predictive solution strategy tags .according to Call the appropriate solver and preprocessor to solve the current system of linear equations.

[0044] In this embodiment, after the predicted solver-preprocessor combination completes the solution of the current linear equation system, its relative residual, convergence state, and solution time are further determined. When the relative residual exceeds a preset residual threshold, the solution time exceeds the maximum solution time limit, or the solver returns to a non-converged state, a fallback solution mechanism is initiated. The fallback solution mechanism includes calling a suboptimal solution strategy or calling a baseline solver preset by the simulation program to complete the solution of the current linear equation system. For samples that trigger fallback, their coefficient matrix is... Right-hand term vector Matrix eigenvectors The actual solver-preprocessor combination used, convergence status, solution time, number of iterations, and relative residuals are saved to the training matrix sample library for subsequent model updates.

[0045] This invention can be implemented through software modules, hardware modules, or a combination of both. The matrix data acquisition module is used to acquire data from geotechnical mechanics simulation programs. and The sample classification module is used to save training samples according to residuals, iteration counts, incremental adjustment states, and fixed interval conditions; the matrix feature extraction module is used to calculate matrix eigenvectors. The candidate combination testing module is used to traverse solver-preprocessor combinations; the label generation module is used to determine the optimal and suboptimal solution strategy labels; and the model training module is used to train the solution strategy selection model. The online prediction module is used to predict the solution strategy during simulation; the solver scheduling module is used to call the corresponding solver and execute the rollback mechanism.

[0046] Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An adaptive optimization solution method for linear equation systems used in geotechnical mechanics simulation, characterized in that, Includes the following steps: S1. Obtain the linear equation system data that actually needs to be solved at the current time step from the data processing process of the geotechnical mechanics simulation program. The linear equation system data includes the coefficient matrix. and the right-hand term vector And determine the vector of the unknown quantity to be solved. ;in, The coefficient matrix is ​​assembled at the current time step. Let be the vector to be solved. It is the right-hand term vector formed by external loads, inertia terms, damping terms, constraint terms, and / or contact action terms; S2. Based on the residual level, error threshold, number of linear iterations, nonlinear or contact state iteration information, time step or load increment adjustment state and fixed calculation step interval in the simulation program data processing process, the linear equation system data is sampled and saved to form a training matrix sample library. S3. Based on the sparse structure, diagonal dominance, symmetry, ill-conditionedness, substructure coupling, right-hand side state, and computational state of linear equation systems, a matrix characteristic index system is constructed to characterize the difficulty of solving linear equation systems, and the coefficient matrix is ​​analyzed. and the right-hand term vector Feature extraction is performed to obtain the matrix eigenvectors; S4. Construct a set of candidate solvers and a set of candidate preprocessors. For each sample in the training matrix sample library, iterate through the combinations of candidate solvers and candidate preprocessors, and record the convergence state, solution time, number of iterations and relative residuals of each combination under the unified residual accuracy requirements and maximum solution time constraints. S5. Based on the solution results recorded in S4, determine the optimal solution strategy label and the suboptimal solution strategy label from the candidate combinations that meet the accuracy requirements. S6. Using the matrix eigenvectors obtained in S3 as input and the optimal solution strategy labels obtained in S5 as output, train the solution strategy selection model. S7. Embed the trained solution strategy selection model into the geotechnical mechanics simulation program, perform online feature extraction on the linear equation system of the new calculation step during the simulation data processing, and predict the corresponding solver-preprocessor combination through the solution strategy selection model. S8. Call the solver-preprocessor combination predicted by S7 to complete the solution of the current linear equation system and return the solution results to the main simulation calculation process; S9. Determine whether the solution result of the current linear equation system in S8 meets the preset residual accuracy requirements and maximum solution time limit; if not, call the suboptimal solution strategy or the preset baseline solver to complete the solution of the current linear equation system, and add the linear equation system data, matrix eigenvectors, actual solver-preprocessor combination, convergence state, solution time, number of iterations and relative residuals to the training matrix sample library for subsequent model updates.

2. The method according to claim 1, characterized in that, In step S2, the sample classification includes at least one of the following: normal convergence matrix, slightly difficult matrix, moderately difficult matrix, difficult matrix, high iteration matrix, time step adjustment matrix, and fixed interval matrix.

3. The method according to claim 2, characterized in that, In step S3, the matrix eigenvector includes at least one of the following: matrix size feature, sparse structure feature, row non-zero element distribution feature, column non-zero element distribution feature, bandwidth feature, diagonal element feature, diagonal dominance feature, off-diagonal coupling feature, symmetry feature, ill-conditioned feature, norm feature, substructure feature, substructure coupling feature, right-hand term vector feature, initial residual feature, solution history feature, and time step state feature, expressed as: ； In the formula, Represents the eigenvectors of a matrix; Indicates the order of a matrix; Indicates the number of active blocks; Representation matrix The number of non-zero elements in the middle; Represents matrix density; Represents the sparsity of a matrix; and These represent the mean, maximum, minimum, and standard deviation of the number of non-zero elements in a row, respectively. and These represent the mean, maximum, and standard deviation of the number of non-zero elements in the column, respectively. and These represent the upper bandwidth, lower bandwidth, and half bandwidth, respectively. , and These represent the mean of the absolute values ​​of the diagonal elements, the maximum absolute value of the diagonal elements, and the minimum absolute value of the non-zero diagonal elements, respectively. and These represent the zero diagonal scale and the near-zero diagonal scale, respectively. and These represent the proportion of diagonally dominant rows and the proportion of strictly diagonally dominant rows, respectively. and These represent the mean and maximum values ​​of the off-diagonal dominance ratio, respectively. This indicates a significant proportion of non-diagonal dominance; Represents matrix symmetry error; Represents an asymmetric, non-zero element proportion; , and Let them represent the matrix 1-norm, infinity norm, and... Norm; This represents the condition number estimate; express Block density; Indicates the number of non-zero 6×6 sub-blocks; Indicates block bandwidth; Indicates the proportion of non-diagonal blocks; Indicates the block coupling strength; Represents the sum of the diagonal block norms; Represents the sum of norms of off-diagonal blocks; , and Let represent the 1-norm, 2-norm, and infinity norm of the right-hand term vector b, respectively; and These represent the mean and standard deviation of the absolute values ​​of the terms on the right-hand side vector, respectively. This indicates that the right-hand side is a non-zero proportion; Indicates the initial residual; This represents the residual from the previous time step; Indicates the number of iterations in the previous time step; Indicates the current time step; A flag variable indicating whether time step adjustment is triggered.

4. The method according to claim 3, characterized in that, In step S3, the matrix size feature includes the matrix order. , non-zero dollar quantity Matrix density sparsity ,in, Coefficient matrix The number of rows or columns, Coefficient matrix The number of non-zero elements and the matrix density sparsity .

5. The method according to claim 4, characterized in that, In step S3, the diagonally dominant feature includes the first absolute value of diagonal elements , No. Sum of absolute values ​​of off-diagonal elements Diagonal dominance ratio off-diagonal dominance ratio ,in: ， ， ， ； Represents the coefficient matrix No. Line number Column elements, For indicator functions, To prevent extremely small positive numbers from being divided by zero.

6. The method according to claim 5, characterized in that, The geotechnical media mechanics simulation data processing employs discontinuous deformation analysis. The method processes the substructure coupling features based on Extract the matrix block structure; In 2D DDA, each active block corresponds to a matrix structure with 6 degrees of freedom, which is extracted, including the number of active blocks. Number of non-zero blocks Block density Non-diagonal block proportions and block coupling strength At least one of them; in, , This indicates the number of 6×6 sub-blocks with a norm greater than zero. , , , Indicates the first The block and the first Each block corresponds to a 6×6 submatrix. express Norm.

7. The method according to claim 6, characterized in that, In step S4, the candidate solver set includes at least one of GMRES, BiCGSTAB, FGMRES, preconditioned conjugate gradient method PCG, minimum residual method MINRES, successive over-relaxation method SOR, and block Jacobi method; the candidate preprocessor set includes at least one of no preprocessing (None), Jacobi preprocessing, Block Jacobi preprocessing, incomplete Cholesky preprocessing (IC), incomplete LU preprocessing (ILU), algebraic multigrid (AMG), and SSOR preprocessing.

8. The method according to claim 7, characterized in that, In step S5, the optimal solution strategy label and the suboptimal solution strategy label are determined, including the following steps: For the For the training matrix sample, calculate the... The solver and the first The relative residuals of each preprocessor combination ; in, For the first The coefficient matrix of each training matrix sample. For the first The right-hand term vector of each training matrix sample. For the first The solver and the first The solution vector obtained by combining several preprocessors; Will satisfy Combinations that are not greater than the preset residual threshold and are successfully solved are sorted in ascending order of solution time. The combination ranked first is labeled as the optimal solution strategy, and the combination ranked second is labeled as the suboptimal solution strategy.

9. A solution system based on the method of any one of claims 1-8, characterized in that, include: The matrix data acquisition module is used to obtain the coefficient matrix and right-hand term vector from the geotechnical mechanics simulation program; The sample classification module is used to save training samples according to residuals, number of iterations, incremental adjustment states, and fixed interval conditions. The matrix feature extraction module is used to calculate the matrix eigenvectors. The candidate combination test module is used to traverse solver-preprocessor combinations; The label generation module is used to determine the labels for the optimal and suboptimal solution strategies; The model training module is used to train the solution strategy selection model; The online prediction module is used to predict the solution strategy during the simulation process; The solver scheduling module is used to call the corresponding solver and execute the rollback mechanism; The sample backflow and model update module is used to add linear equation samples that do not meet the preset accuracy requirements, exceed the maximum solution time, or are newly added, along with their actual solution results, to the training matrix sample library for subsequent model updates.

Images