A neural network prediction method and system for hard-constrained physical information of rock mass deformation over time
By using a hard-constrained physical information neural network architecture and a robust loss function system, the problems of insufficient boundary condition satisfaction and numerical oscillations caused by higher-order derivatives in rock aging deformation are solved, achieving high-precision and stable prediction of rock aging deformation.
Patent Information
- Application Number
- CN202610481881.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2026-04-13
- Publication Date
- 2026-11-17
- Estimated Expiration
- 2046-04-13
AI Technical Summary
Existing physical information neural networks suffer from problems such as insufficient accuracy in satisfying boundary conditions and long-term extrapolation failure due to the accumulation of boundary errors over time when processing the time-dependent deformation of rocks. Furthermore, the numerical oscillations and gradient explosion problems caused by higher-order derivatives are difficult to solve.
We adopt a hard-constrained physical information neural network architecture, enforce the Dirichlet boundary conditions by using a geometric distance factor, and construct a robust loss function system using Huber loss function and Smooth-L1 loss function. Combined with gradient pruning and dynamic learning rate scheduling strategies, we achieve coordinated convergence of data-driven and physical-driven approaches.
It effectively eliminates the accumulation of boundary errors, suppresses gradient explosion, ensures the accuracy and stability of the model in long-term prediction, and improves the accuracy and computational efficiency of rock age deformation prediction.
Smart Images

Figure CN122616636B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of rock mechanics technology, specifically to a method and system for predicting the time-dependent deformation of rock masses using a hard-constrained physical information neural network. Background Technology
[0002] The rheological properties of rocks are a key factor affecting the long-term stability of engineering projects. Time-dependent deformation of rocks often leads to rock mass instability and even structural failure. Therefore, establishing high-precision prediction models is of great significance for assessing the long-term stability of engineering projects. In recent years, data-driven methods, such as deep neural networks (DNNs) and long short-term memory networks (LSTMs), have shown significant advantages in soil and rock displacement prediction due to their powerful nonlinear mapping capabilities. However, purely data-driven models are essentially "black box" probabilistic statistical models, lacking explanations of physical mechanisms; moreover, their performance is highly dependent on the data scale. In the typical "small sample, high noise" conditions of geotechnical engineering, they are prone to overfitting and may even produce prediction results that violate basic physical conservation laws.
[0003] To overcome the aforementioned bottlenecks, Physics-Informed Neural Networks (PINN) has achieved deep integration of data and physical laws by embedding physical control equations as regularization terms into the loss function of the neural network, and has made progress in fields such as fluid mechanics and thermodynamics. In the field of geotechnical engineering, PINN also shows great promise: Previous studies have incorporated the Terzaghi consolidation equation into the PINN model to address the consolidation problem of layered foundations, with the L2 norm error between the simulated and measured consolidation coefficients controlled within 0.005; the Richards equation has been incorporated into PINN to study soil moisture content under homogeneous conditions, improving prediction accuracy by 17% compared to traditional numerical software; based on the constitutive equations of elasticity, PINN has been used to predict surface settlement induced by shield tunnel construction and to invert Poisson's ratio and elastic modulus parameters, with simulation results showing good agreement with measured values; the integration of penetration mechanics models into PINN for consolidation coefficient inversion has resulted in an L2 norm error of only 0.019 between the simulated values and pore pressure static cone penetration test data; and the PINN model coupled with the Navier-Cauchy equations has been used to invert rock mechanical parameters of shale oil reservoirs. PINN's prediction determination coefficients (R²) all exceeded 0.95, representing improvements of approximately 10%, 5%, and 9% respectively compared to traditional machine learning models such as random forest, XGBoost, and artificial neural networks. A pile foundation m-method PINN analysis model considering load and pile-soil parameters was proposed. Through hard constraints and dimensionless parameterization, training was accelerated, achieving a computation time of only 0.0019 seconds for reliability analysis of millions of samples, significantly improving computational efficiency compared to the finite element method. Other research proposed a probabilistic physical information-guided landslide susceptibility assessment model, which improved the average AUC in spatial cross-validation by 13.6% compared to the purely data-driven model, while reducing the scientific inconsistency index by 88.6%. Furthermore, a PINN method constructing a coupled hybrid phase-field fracture model was used to simulate rock tensile-shear failure, and the joint optimization algorithm employed improved convergence accuracy by 16.96% compared to traditional methods.
[0004] Although PINN has shown great promise in the aforementioned geotechnical static inversion and multi-field coupled simulations, research on PINN for the time-dependent deformation characteristics of rocks remains relatively scarce. The main reason for this is the computational complexity and non-convex optimization difficulties arising from the significant time-dependent rheological characteristics of rocks. Specifically, using PINN to predict the time-dependent deformation of rocks mainly faces the following challenges: First, the standard PINN uses soft constraints to handle initial and boundary conditions, causing the residuals of the physical equations and the residuals of the boundary conditions to compete with each other during training. This makes it difficult for the model to accurately satisfy the boundary constraints, ultimately resulting in low overall prediction accuracy. In particular, for rock rheology, a typical "long-duration" physical process, the small residual deviations at the boundary under the soft constraint mechanism will accumulate over time and be amplified by higher-order derivatives in long-term predictions, which can easily destroy physical consistency and lead to long-term extrapolation failure.
[0005] Second, due to the order-of-magnitude differences in multi-objective loss functions, gradient imbalance or anomalies are highly likely to occur during training, causing the model to get stuck in local optima. The rock rheological constitutive equation is a rigid differential equation containing high-order time derivatives. In the early stages of training, the random initialization of neural network parameters leads to an extremely non-smooth output field in the spatiotemporal domain. High-frequency noise, amplified by high-order differential operators, generates singular physical residuals of enormous magnitude. If a traditional mean squared error loss function is used, these outliers will generate huge quadratic penalty gradients, easily triggering gradient explosion and thus compromising the stability of the optimization process.
[0006] Therefore, constructing an architecture with strict mathematical boundary constraints and introducing more robust optimization strategies are key to improving PINN's ability to predict rock deformation over time. Summary of the Invention
[0007] To address the aforementioned technical problems, this invention proposes a neural network prediction method and system for time-dependent deformation of rock masses based on hard-constrained physical information. The technical problems it solves include: (1) In view of the technical problems of insufficient accuracy in satisfying boundary conditions under soft constraint mechanism and long-term extrapolation failure caused by the accumulation of boundary error over time in existing physical information neural networks, this invention constructs a hard constraint network architecture based on geometric distance factor, which forces the Dirichlet boundary conditions to be satisfied from a mathematical structure perspective, so that the satisfaction of boundary conditions is completely independent of the optimization process of network parameters, thus fundamentally eliminating the boundary effect.
[0008] (2) To address the numerical oscillations and gradient explosion caused by higher-order derivatives in the rock rheological constitutive equation, this invention replaces the traditional mean square error loss function with Huber loss function and Smooth-L1 loss function, and constructs a dual robust loss function system, which effectively suppresses abnormal gradients in the process of solving rigid equations.
[0009] (3) To address the gradient competition problem between data loss and physical loss in multi-objective loss functions, this invention introduces a dynamic weight balancing algorithm based on gradient norm, combined with gradient pruning mechanism and linear preheating-cosine annealing learning rate scheduling strategy, to achieve coordinated convergence of data-driven and physical-driven mechanisms throughout the training process.
[0010] To address the aforementioned technical problems, this invention provides a neural network prediction method for time-dependent deformation of rock mass based on hard-constrained physical information, comprising the following steps: Step S1: Construct a deep neural network with normalized spatiotemporal coordinates as input and displacement vector as output; Step S2: Multiply each component of the displacement vector by the corresponding geometric distance factor to obtain the predicted displacement field. The geometric distance factor is zero on the target boundary, so that the predicted displacement field satisfies the Dirichlet boundary condition on the target boundary. Step S3: Calculate the partial derivatives of the predicted displacement field and substitute them into the rock rheological constitutive equation to calculate the physical residuals; Step S4: The deviation between the predicted displacement field and the monitoring data, as well as the physical residual, are weighted and summed using a robust loss function to construct a total loss function; Step S5: Minimize the total loss function to train the deep neural network, and use the predicted displacement field as the prediction result after convergence.
[0011] Preferably, in step S2, the geometric distance factor is defined in the solution domain. Generalized distance function on The generalized distance function Satisfy: When the spatial coordinates point Located in the solution domain boundary When, When spatial coordinates point Located in the solution domain Inside, .
[0012] Preferably, for the solution domain of a cylindrical specimen under axisymmetric conditions, the target boundary includes the axis of symmetry and the bottom end face, and the generalized distance function is taken as the ratio of the radial coordinate to the specimen width and the ratio of the axial coordinate to the specimen height, respectively; the radial predicted displacement of the predicted displacement field... and axial predicted displacement They are defined as follows: ; ; in, Radial coordinates, For axial coordinates, For time, The width of the sample. The height of the sample. and These are the radial and axial displacement components output by the deep neural network, respectively. This is the set of trainable parameters for a deep neural network.
[0013] Preferably, the rock rheological constitutive equation in step S3 is a three-dimensional deviatoric stress-deviatoric strain tensor differential equation of the generalized Kelvin model, expressed as: ; in, These are components of the deviatoric stress tensor. These are the components of the deviatoric strain tensor. The elastic shear modulus of the Hooke body. The elastic shear modulus of the Kelvin body. The shear viscosity coefficient of Kelvin volume. express directional surface along The deviatoric stress component acting in the direction of the stress. express direction and Deviatoric strain components between directions, subscript and Values and ,in Radial, For the axial direction.
[0014] Preferably, the physical residual in step S3 includes constitutive equation residuals. Stress boundary residuals Poisson effect residuals Three components; The constitutive equation residual The expression is: ; The stress boundary residual The expression is: ; The Poisson effect residual The expression is: ; In the formula, Predict stress at the boundary. For the true boundary stress, The radial strain value is calculated based on the predicted displacement field. The axial strain value is calculated based on the predicted displacement field. It is Poisson's ratio.
[0015] Preferably, the total loss function in step S4 Represented as: ; in, This is the set of trainable parameters for a deep neural network. The data loss term is calculated using the Huber loss function based on the deviation between the predicted displacement field and the monitoring data. The physical loss term is calculated using the Smooth-L1 loss function on the physical residuals. and These are the weighting coefficients for the data loss term and the physical loss term, respectively.
[0016] Preferably, the weighting coefficient and Adaptive adjustment is performed using a gradient norm-based dynamic adjustment algorithm: the data loss term is recorded in real time during training. and the physical loss term Each historical decline rate is used to adaptively adjust the weight coefficient of the physical loss term based on the ratio of the historical decline rates at preset fixed iteration intervals. This allows the data loss term and the physical loss term to converge collaboratively throughout the training process.
[0017] Preferably, the collocation points used in step S3 to calculate the physical residual are generated using a spatiotemporal non-uniform sampling strategy, specifically: spatial domain collocation points are generated on the sample cross-section using Latin hypercube sampling technology; time domain collocation points are generated using a mixed logarithmic distribution sampling strategy.
[0018] Preferably, the deep neural network includes an input layer, multiple fully connected hidden layers, and an output layer; the input layer receives three-dimensional normalized spatiotemporal coordinates composed of radial coordinates, axial coordinates, and time; the neuron output of each fully connected hidden layer undergoes a nonlinear transformation through a hyperbolic tangent activation function, the... Layer output vector Represented as: ; in, and The first Layer weight matrix and bias vector, The activation function is hyperbolic tangent; the weight parameters of the deep neural network are initialized using the Xavier normal distribution strategy. Step S5, when minimizing the total loss function, also includes the following training strategy: Gradient clipping: Before each parameter update, the gradient norm of the total loss function with respect to the deep neural network is calculated, and the gradient is clipped when the gradient norm exceeds a preset clipping threshold. Dynamic learning rate scheduling: A learning rate scheduling strategy combining a linear warm-up phase and a cosine annealing phase is adopted. In the linear warm-up phase, the learning rate increases linearly from zero to the peak learning rate. In the cosine annealing phase, the learning rate decays from the peak learning rate to the minimum learning rate according to the cosine function curve.
[0019] This invention also provides a hard-constrained physical information neural network prediction system for time-dependent deformation of rock mass, applicable to the above-mentioned method, including: The neural network building module is used to build deep neural networks that take normalized spatiotemporal coordinates as input and displacement vectors as output. The hard constraint module is used to multiply each component of the displacement vector by the corresponding geometric distance factor to obtain the predicted displacement field. The geometric distance factor is zero on the target boundary, so that the predicted displacement field satisfies the Dirichlet boundary condition on the target boundary. An automatic differential physics engine module is used to calculate the partial derivatives of the predicted displacement field and substitute them into the rock rheological constitutive equation to calculate the physical residuals; The loss function construction module is used to construct a total loss function by weighted summation of the deviation between the predicted displacement field and the monitoring data and the physical residual using a robust loss function. An adaptive optimization module is used to minimize the total loss function to train the deep neural network, and after convergence, the predicted displacement field is used as the prediction result.
[0020] The beneficial effects of the present invention include at least the following: (1) This invention forces the Dirichlet boundary conditions to be met mathematically by multiplying the network output by a geometric distance factor that is always zero on the target boundary, making the satisfaction of the boundary conditions completely independent of the optimization process of the network parameters. This fundamentally eliminates the systematic error accumulation caused by boundary drift under the soft constraint mechanism. At the same time, this hard constraint architecture is not limited to regular geometries. For complex geotechnical engineering solution domains of arbitrary shapes, a generalized distance function that satisfies the conditions can be constructed, which has wide engineering adaptability.
[0021] (2) This invention employs a dual robust loss function system constructed using the Huber loss function and the Smooth-L1 loss function, effectively solving the numerical rigidity problem caused by higher-order derivatives in the generalized Kelvin constitutive equation. In the early stages of training, the random initialization of network parameters leads to outliers of extremely large magnitudes in the physical residuals. Traditional mean squared error loss functions amplify these outliers into huge quadratic penalty gradients, causing gradient explosion. The dual robust loss function effectively suppresses the destructive impact of outlier gradients on the optimization process by switching to a linear penalty form in the large error region, eliminating the destructive gradient bounce spikes that occur in hard-constrained models when using the mean squared error loss function.
[0022] (3) This invention solves the gradient competition dilemma caused by the difference in magnitude between multiple objective loss functions by real-time monitoring of the historical decline rate of the data loss term and the physical loss term and adaptively adjusting the weight coefficients, avoids the blindness of manual parameter tuning, and ensures the coordinated convergence of the data-driven and physical-driven mechanisms throughout the training process.
[0023] (4) The present invention makes the density of physical points dense in the initial stage of the time domain and sparse in the later stage. This distribution feature is highly consistent with the physical feature that the deformation rate decreases exponentially with time in the generalized Kelvin model. It focuses limited computing resources on the key areas where physical changes are most intense, and improves the model’s accuracy in capturing transient features of rock rheology.
[0024] (5) This invention uses a dynamic learning rate scheduling strategy that combines gradient pruning mechanism with linear warm-up-cosine annealing to enable the network to start smoothly in the early stage of training and avoid divergence due to excessive initial step size, maintain a high learning rate in the middle stage of training to quickly reduce the loss value, and smoothly decay in the later stage to finely approximate the optimal solution. Attached Figure Description
[0025] Figure 1 This is a schematic diagram of the overall architecture of the method in an embodiment of the present invention; Figure 2 This is a schematic diagram of a deep neural network structure according to an embodiment of the present invention; Figure 3 This is a schematic diagram of the rheological curve of chlorite schist according to an embodiment of the present invention; Figure 4 This is a schematic diagram comparing the rheological curve fitting results of embodiments of the present invention; Figure 5 This is a schematic diagram of a spatiotemporal non-uniform sampling strategy adapted to physical characteristics according to an embodiment of the present invention; Figure 6 The above are relative error distribution cloud maps of four schemes in the embodiments of the present invention; Figure 7 This is a schematic diagram of the iterative convergence curve of the M2 model in an embodiment of the present invention; Figure 8 This is a schematic diagram of the iterative convergence curve of the M3 model in an embodiment of the present invention; Figure 9 This is a schematic diagram comparing experimental and predicted rheological curves in an embodiment of the present invention. Detailed Implementation
[0026] 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 protection scope of the present invention.
[0027] This invention proposes a rock deformation prediction method based on a geometrically hard-constrained physical information neural network (HC-PINN). It uses a deep neural network as an approximation of the displacement field. By embedding a hard-constraint module based on a geometric distance factor at the network output, boundary conditions are mathematically enforced. Automatic differentiation techniques are then used to substitute the predicted displacement field into the rock rheological constitutive equation to calculate the physical residuals. Finally, the network parameters are trained by minimizing the total loss function, which includes both data and physical losses. Figure 1 As shown, the model mainly consists of four parts: a deep neural network, a hard constraint module, an automatic differential physics engine, and an adaptive optimization module. The technical implementation of each module is explained in detail below.
[0028] Step S1: Construct a deep neural network with normalized spatiotemporal coordinates as input and displacement vector as output.
[0029] This invention employs a deep neural network as an approximate approximator for the displacement field. For example... Figure 2 As shown, the network architecture consists of an input layer, four fully connected hidden layers (64 neurons each), and an output layer. The network's input vector is a normalized three-dimensional spatiotemporal coordinate, including radial coordinates. Axial coordinates and time The three components are represented as The network output is a displacement vector, including radial displacement components. and axial displacement components Information transfer between adjacent layers of a network follows a nonlinear mapping rule. The output vector of a neuron in a layer can be represented as: ; in, and They represent the first Layer weight matrix and bias vector, This represents a nonlinear activation function. This invention selects the hyperbolic tangent function as the activation function, and its output range is... It possesses smooth differentiability, which is beneficial for the stable propagation of gradients during automatic differentiation. The network weights are initialized using a Xavier normal distribution strategy, effectively preventing the vanishing gradient phenomenon during deep network training.
[0030] Step S2: Multiply each component of the displacement vector by the corresponding geometric distance factor to obtain the predicted displacement field. The geometric distance factor is zero on the target boundary, so that the predicted displacement field satisfies the Dirichlet boundary condition on the target boundary.
[0031] Standard physical information neural networks (PENs) use soft constraints to handle boundary conditions, adding the boundary condition residuals as an additional loss term to the total loss function and iteratively approximating the zero value through the optimizer. However, this approach cannot mathematically guarantee that the boundary conditions are precisely satisfied. Especially in long-duration physical processes like rock rheology, tiny residual deviations at the boundaries accumulate over time, eventually compromising physical consistency.
[0032] To address the aforementioned problems, this invention designs a hard-constraint architecture based on the geometric distance factor. The core idea is to introduce a multiplication module at the network output, multiplying the original output of the deep neural network by the geometric distance factor component by component. The geometric distance factor is constructed based on the principle that it is always zero on the target boundary where zero displacement constraints need to be applied, and takes a positive value within the solution domain. Since the product of any real number and zero is zero, regardless of the network output value, the predicted displacement field after passing through the hard-constraint module must be zero on the target boundary, thus mathematically forcing the satisfaction of the Dirichlet boundary conditions. This satisfaction is unconditional and completely independent of the specific values of the network parameters.
[0033] In this embodiment, the cylindrical specimen is constrained under axisymmetric conditions. The target boundary includes the axis of symmetry (the position where the radial coordinate is zero) and the bottom end face (the position where the axial coordinate is zero). The corresponding geometric distance factors are taken as the ratio of the radial coordinate to the specimen width and the ratio of the axial coordinate to the specimen height, respectively. Accordingly, the predicted displacement field after processing by the hard constraint module is defined as: ; ; in, Radial coordinates, For axial coordinates, For time, The width of the sample. The height of the sample. and These are the radial and axial displacement components output by the deep neural network, respectively. This represents the complete set of trainable parameters for the deep neural network. At the axis of symmetry, the radial coordinates... The geometric distance factor is zero. The value is zero, making the radial predicted displacement zero. The coordinates must be zero at the axis of symmetry. Similarly, at the bottom end face, the axial coordinates... Zero, axial predicted displacement It must be zero at the bottom end face.
[0034] It should be noted that although this embodiment uses a regular cylindrical specimen as an example to construct a simple linear distance factor, this hard constraint framework is inherently highly adaptable and not limited to regular geometries. A generalized distance function can be constructed for complex geotechnical engineering solution domains Ω of arbitrary shapes. As long as the function satisfies the condition on the boundary of the solution domain And within the domain Given the given conditions, the above hard constraint form can be extended to any irregular boundary condition.
[0035] Step S3: Calculate the partial derivatives of the predicted displacement field and substitute them into the rock rheological constitutive equation to calculate the physical residuals.
[0036] The core task of this step is to use automatic differentiation technology to calculate the partial derivatives of each order of the predicted displacement field obtained in step S2, construct the physical quantities required for the rock rheological constitutive equation based on the partial derivatives, and then substitute them into the constitutive equation to calculate the physical residuals. This section introduces in turn the selection of the constitutive equation, the construction process of the physical quantities, the definition of the physical residuals, and the sampling strategy for spatiotemporal collocation.
[0037] The age-dependent deformation of rock in this invention follows the viscoelastic-rheological behavior described by the generalized Kelvin model. The generalized Kelvin model consists of a Hooke elastic body and a Kelvin viscoelastic body connected in series, accurately simulating the instantaneous elastic deformation and decaying rheological deformation of rock under constant stress. Its three-dimensional tensor differential form is as follows: ; in, For the components of the deviatoric stress tensor, representing directional surface along The deviatoric stress component acting in the direction; For the components of the partial strain tensor, representing direction and Deviatoric strain components between directions; subscript and The range of values is and ,in Radial, For the axial direction; The elastic shear modulus of the Hooke body. The elastic shear modulus of the Kelvin body. is the shear viscosity coefficient of Kelvin.
[0038] The constitutive equations described above involve the deviatoric stress tensor, the deviatoric strain tensor, and their time derivatives. These physical quantities need to be constructed step-by-step from the predicted displacement field output in step S2 using an automatic differentiation technique. The automatic differentiation technique is a precise differentiation calculation method based on the chain rule. It utilizes the computational graph of a deep learning framework to record all computational operations during the forward propagation process, and then traces back through the computational graph to accurately obtain the partial derivatives of the output with respect to the input of any order. The specific construction process is as follows: First, the strain tensor components are obtained by taking the partial derivatives of the predicted displacement field with respect to each spatial coordinate. Then, the partial strain tensor components are obtained by subtracting the volumetric strain spherical tensor components from the strain tensor components. Next, the stress tensor components are calculated from the strain tensor components based on the elastic constitutive relation, and the partial stress tensor components are obtained by subtracting the mean stress spherical tensor components. Finally, the partial derivatives of the partial strain and partial stress tensor components with respect to time are taken respectively. It is particularly important to note that the calculation of the partial strain rate essentially involves the mixed second derivatives of the predicted displacement field with respect to space and time, which is the fundamental source of the numerical stability challenge in solving rigid equations.
[0039] After substituting the physical quantities constructed above into the generalized Kelvin constitutive equation, this invention defines three physical residual components: constitutive equation residual, stress boundary residual, and Poisson effect residual.
[0040] constitutive equation residuals Represented as: ; Stress boundary residual Represented as: ; in, Predict the stress value for the boundary. This represents the true stress value at the boundary.
[0041] Poisson effect residual Represented as: ; in, The radial normal strain is calculated based on the predicted displacement field. The axial normal strain is calculated based on the predicted displacement field. It is Poisson's ratio.
[0042] In the training process of a physical information neural network, the quality and distribution of samples directly determine the training effect. This invention designs a hybrid sampling strategy to address the time-dependent deformation characteristics of rocks, aiming to reallocate computational resources to improve the network's ability to capture physical field features.
[0043] In this embodiment, for the two-dimensional solution domain, Latin Hypercube Sampling (LHS) is used to generate spatial collocation points. LHS sampling is performed on the longitudinal section of the sample. This technique ensures uniform coverage of sampling points in the high-dimensional space by dividing the range of values for each dimension into several layers and randomly sampling within each layer, thus avoiding potential clustering or blank areas in random sampling.
[0044] For the time domain, considering the non-uniform decay of rock deformation rates over time, a hybrid logarithmic distribution sampling strategy was designed, drawing inspiration from non-uniform sampling in high-gradient problems. This strategy results in an exponential decay in collocation density over time—dense collocations are placed during the initial stage of loading when deformation rates are high and physical field changes are drastic, while the density gradually decreases in the later stage when deformation tends to stabilize. This distribution characteristic is highly consistent with the mathematical features of the exponential decay term in the constitutive equation of the generalized Kelvin model. Compared with the traditional uniform sampling strategy, the hybrid logarithmic distribution sampling has a significantly higher collocation density in the transient response stage, ensuring that the model can focus computational resources on the key regions where physical changes are most drastic, achieving high-precision prediction across the entire time domain.
[0045] Step S4: Use a robust loss function to weight and sum the deviations between the predicted displacement field and the monitoring data, as well as the physical residuals, to construct the total loss function.
[0046] This step constructs a total loss function that includes data loss terms and physical loss terms, and uses a robust loss function to replace the traditional mean square error function to address the numerical challenges posed by high-order rigid differential equations.
[0047] In this embodiment, a total loss function including data loss term and physical loss term is constructed, and a robust loss function is used to replace the traditional mean square error function to deal with the numerical problems caused by high-order rigid differential equations.
[0048] Directly using automatic differential (AD) techniques to calculate residuals in the generalized Kelvin constitutive equations faces significant numerical stiffness challenges. This is because the equations include strain rate terms. This essentially involves displacement fields. The second derivative with respect to both space and time. In the early stages of training, the random initialization of neural network parameters leads to an extremely non-smooth output field in the spatiotemporal domain. This high-frequency noise, amplified by higher-order differential operators, produces singular physical residuals of an order of magnitude larger. If a traditional mean squared error loss function is used, these outliers will generate huge quadratic penalty gradients, which can easily trigger gradient explosion and thus destroy the stability of the optimization process.
[0049] Therefore, this invention employs the Huber loss function to calculate the data loss term and the Smooth-L1 loss function to calculate the physical loss term, replacing the traditional MSE norm. The Huber loss function is calculated in squared error form when the prediction bias is small, maintaining the sensitivity of the mean squared error in the small error region; when the prediction bias is large, it switches to linear form, effectively suppressing the excessive influence of outliers on the gradient. The Smooth-L1 loss function has similar robustness, calculating in squared form when the residual value is small to ensure smoothness, and calculating in linear form when the residual value is large to reduce sensitivity to outliers.
[0050] In this embodiment, the total loss function Defined as the weighted sum of data loss and physical loss: ; in, and These are the weighting coefficients for the data loss term and the physical loss term, respectively. For data loss items, This represents the physical loss term. Expanding the above equation into a specific summation calculation: ; in, These are the trainable parameters of a deep neural network. The number of data samples, The number of physical points, Here is the Huber loss function. For the model in the first Predicted displacement values at each data sample point For the corresponding number The true displacement value of each data sample point The Smooth-L1 loss function is used. For the first Internal static weighting coefficients for each physical residual component For the first Each physical residual component and The first Spatial and temporal coordinates of each physical point, subscript The range of values is These correspond to the three components: constitutive equation residual, stress boundary residual, and Poisson effect residual, respectively. Let denote the vector norm, where subscript 2 in the data loss term indicates the L2 norm, and subscript 1 in the physical loss term indicates the L1 norm.
[0051] In the training of a physical information neural network, the data loss term With physical loss item Often, there are differences in magnitude or inconsistencies in convergence speed. Direct summation can lead to the optimization process being dominated by terms with larger values, resulting in a multi-task competition dilemma. To address this ill-conditioned gradient problem, this embodiment of the invention introduces a dynamic adjustment algorithm based on the gradient norm to balance the weight coefficients in the total loss function. and By recording and evaluating the historical decline rates of the data loss term and the physical loss term in real time, the weight coefficient of the physical loss term is adaptively adjusted every fixed iteration. This strategy assigns greater weight to loss terms that have slower training rates, ensuring that both data-driven and physical-driven mechanisms converge synergistically throughout the training process. This effectively avoids the pitfalls of manual parameter tuning.
[0052] Step S5: Minimize the total loss function to train the deep neural network, and use the predicted displacement field as the prediction result after convergence.
[0053] To find the global optimal solution in the parameter space and avoid numerical oscillations caused by high-order rigid differential equations, this invention proposes a combined optimization strategy that includes random mini-batch training, gradient pruning, and dynamic learning rate scheduling.
[0054] This invention employs the Adam optimizer for parameter updates. To address issues such as memory overflow and getting trapped in local minima that may occur during full-data training, a mini-batch stochastic gradient descent strategy is used to enhance the model's generalization ability.
[0055] To address the gradient explosion problem that may be caused by rigid equations, this invention introduces a gradient pruning mechanism. Before each parameter update, the gradient norm of the total loss function with respect to all parameters of the deep neural network is calculated. When the gradient norm exceeds a preset pruning threshold, the gradient is pruned to prevent excessively large single-step parameter updates from causing training divergence. In this embodiment, the pruning threshold is set to 0.5.
[0056] To balance exploratory capabilities in the early stages of training with convergence accuracy in the later stages, this invention designs a dynamic learning rate scheduling strategy combining linear warm-up and cosine annealing. In the early stages of model training, a warm-up phase is set up where the learning rate linearly increases from zero to an initial peak learning rate, allowing the network to start smoothly in an initial random state and avoiding training divergence caused by excessively large initial step sizes. After warm-up, the learning rate gradually decays to its minimum value following a cosine function curve. This strategy ensures that the learning rate remains at a high level in the middle stages of training to rapidly reduce the loss value, and then decays smoothly in the later stages to achieve a fine approximation of the optimal solution. Finally, combined with a dynamic weight strategy, it achieves high-precision convergence of the model.
[0057] To verify the effectiveness of the method of the present invention, a uniaxial compression rheological test was conducted on a chlorite schist sample with a dip angle of 0°. The rock sample was processed into a standard cylindrical sample with a diameter of 50 mm × 100 mm. The test was conducted using a high confining pressure electro-hydraulic servo rock triaxial rheological testing machine, and the ambient temperature was kept constant at 25°C.
[0058] Based on a pre-conducted conventional uniaxial compression test, the Poisson's ratio of the specimen with a bedding angle of 0° was measured. The value is 0.272, and the uniaxial compressive strength is 120 MPa. Based on this strength, a single-sample multi-stage loading test method was adopted, and six stress levels were set as shown in Table 1.
[0059] Table 1 The instantaneous strain, rheological strain, loading time, and rheological and instantaneous strain of the specimens under various stress levels are shown in Table 2 and... Figure 3 As shown.
[0060] Table 2 The rheological curves of chlorite schist exhibit distinct staged characteristics. The samples show significant decaying rheology, stabilizing from the third loading stage with small deformation increments, and failing at the final loading stage. At the first stress level (43.7 MPa), the samples demonstrate a clear decaying rheological process, with rheological strain accounting for as much as 22.9%. Compared to the steady-state rheological response at subsequent loading levels, the first-stage rheological curve exhibits significant viscoelastic characteristics. Therefore, using the rheological data from the first loading stage as the research object for HC-PINN is more effective in verifying the model's predictive ability and its ability to solve the rock rheological equations.
[0061] Based on the Boltzmann superposition principle, a method combining the least squares method and FLAC3d numerical simulation was used to identify rheological parameters. The parameter identification results are shown in Table 3. Figure 4 As shown, the experimental data and the fitted curve have a very high degree of consistency, and the identified rheological parameters can accurately reflect the viscoelastic behavior of the rock under the current stress level.
[0062] Table 3 Based on the aforementioned hybrid sampling strategy and rheological data under the first stress level, a hybrid training dataset containing data sample set and physical sample set was constructed.
[0063] The data sample set was constructed to simulate limited, discrete monitoring data in practice. Spatially, it followed the principle of sparse boundary monitoring, randomly sampling only 12 spatial feature points along the geometric boundaries of the specimen. In terms of time series, it simulated a short-term monitoring scenario, capturing only the first 10% of the rheological test duration (i.e., 30 hours) and employing a variable-interval sampling strategy, performing high-frequency sampling at the initial loading stage and gradually decreasing the sampling frequency thereafter. Displacement label values were calculated based on the analytical solution of the rheological constitutive equations of the generalized Kelvin model. ; ; ; ; in, For the applied constant axial stress, Bulk modulus The value is Poisson's ratio. A total of 432 spatiotemporal data samples were ultimately generated for calculating the data loss term.
[0064] The physical sample set aims to compensate for the lack of data samples within the specimen and in the long time domain, utilizing the residuals of the physical equations for meshless self-supervised learning. The physical solution domain is defined as follows: The two-dimensional spatial domain is The unit is m, corresponding to the longitudinal section of a Φ50mm×100mm sample, and the time domain is... The unit is h. A total of 10,000 physical collocation points were generated in the full-spacetime solution domain, and their spatial and temporal distributions are as follows: Figure 5 As shown, Figure 5 In this context, (a) represents the spatial sampling strategy. Figure 5 (b) in the text represents the time sampling strategy.
[0065] To systematically evaluate the effectiveness of the HC-PINN architecture in the embodiments of this invention, four progressive computation schemes were designed, following the evolutionary logic of physical mechanism introduction—boundary architecture improvement—synthetic strategy optimization: Scheme M0 is a purely data-driven deep neural network without any physical constraints, employing a mean squared error loss function and an Adam optimizer (fixed learning rate). Scheme M1 is a standard physically-informed neural network, using soft constraints to handle boundary conditions, employing a mean squared error loss function, Latin hypercube spatial sampling, and an Adam optimizer (fixed learning rate). Scheme M2 changes the boundary handling method from soft constraints to hard constraints, with the remaining configurations being the same as M1. Scheme M3 is the complete HC-PINN model proposed in this invention, integrating all improved strategies on the basis of hard constraints, including a dual robust loss function, spatial Latin hypercube combined with temporal mixed logarithmic sampling, gradient pruning, dynamic weight balancing, and linear preheating plus cosine annealing learning rate scheduling.
[0066] All four schemes use the same underlying neural network architecture (4 hidden layers, 64 neurons per layer, hyperbolic tangent activation function), and the maximum number of iterations is set to 30,000.
[0067] The specific configuration strategies and core mechanism differences of each scheme are summarized in Table 4.
[0068] Table 4 To quantitatively evaluate the predictive performance of each model, relative error is used as the core evaluation metric. For any given model... The relative error of each spatiotemporal monitoring point is defined as: ; in, The model represents the first time. Predicted displacement values for each monitoring point This represents the baseline value of the analytical solution of the generalized Kelvin model.
[0069] Figure 6 The diagram shows the relative error distribution contour plots of the four schemes on the validation set. To clearly illustrate the error distribution characteristics of each model, different subplots use independent color ranges. Figure 6 In (b), the bottommost line appears to have zero error, but this is actually because soft constraints cannot mathematically make the predicted value... , and the benchmark value This causes the relative error to degenerate into the absolute error form due to the singularity of the denominator, thus forming a pseudo-low error line at the bottom.
[0070] (1) Limitations of the data-driven model: such as Figure 6 As shown in (a), the purely data-driven M0 model exhibits significant prediction bias across the entire domain, with a maximum relative error as high as 295.3%. In the axial coordinate... Near the bottom region of 0, a deep red high error band appears. Due to the lack of constraints from physical equations and the sparse training data, the neural network struggles to learn the correct mechanical gradient in the boundary region where no data is covered, leading to severe numerical divergence.
[0071] (2) Boundary drift and singularity amplification of soft constraints: such as Figure 6 As shown in (b), the M1 model with soft constraints introduced by the physical equations still has a significant red high error zone at the bottom of the sample, with a maximum error of about 95.0%. This is not a simple boundary intercept error, but because under the soft constraint mechanism, the model lacks strict mathematical structural constraints in the low-value region near the boundary. Although the model may approach the zero value at the boundary point through optimization, it is difficult to accurately capture the small deformation gradient near the boundary. In the bottom region where the actual displacement is very small, the small prediction deviation is amplified by the relative error, thus forming the systematic high error band shown in the figure.
[0072] (3) A qualitative change in the prediction accuracy of hard constraints: such as Figure 6 As shown in (c), the systematic error band at the bottom has completely disappeared, indicating that the hard constraint architecture mathematically locks the Dirichlet boundary conditions. M2 has successfully reduced the maximum error from 95% of M1 by two orders of magnitude to below 0.2%, verifying the advantage of the hard constraint architecture in eliminating boundary effects.
[0073] Figure 7 and Figure 8 The training iteration convergence curves of models M2 and M3 are shown. The loss curve of M2 exhibits sawtooth oscillations during its descent. This is because, with the MSE loss function and a fixed learning rate, the optimizer struggles to adapt to large gradient changes, leading to repeated gradient jumps and highly unstable convergence. Furthermore, around iteration 21,000, the loss curve shows an extreme rebound peak, caused by a sudden surge in gradient values due to local higher-order derivatives of the rigid differential equation. In contrast, the M3 model, which integrates a dual robust loss function, gradient clipping, and dynamic weighting strategies, demonstrates excellent convergence characteristics. Figure 8 As shown, although the curve exhibits a certain degree of sawtooth fluctuation, this is mainly due to the gradient noise introduced by the Mini-batch stochastic gradient descent strategy. While this randomness leads to local oscillations, it effectively helps the model escape local minima and enhances its global optimization ability. Thanks to the tolerance of the dual robust loss function for gradient outliers and the mandatory constraint of the parameter update step size by the gradient pruning mechanism, M3 effectively suppresses the destructive gradient explosion that occurs in M2.
[0074] Overall, the convergence curve of M3 shows a clear and monotonous downward trend, with the data loss remaining stable at an extremely low 10%. -9The order of magnitude, physical loss converges to 10. -5 This indicates that the comprehensive robust strategy successfully transformed the anomalous oscillations of the rigid differential equation into a controllable stochastic exploration, achieving a balance between training accuracy and stability.
[0075] Figure 9 The rheological prediction curves of M3 and M0 are presented, along with the theoretical solutions and experimental data of the generalized Kelvin model. Although the purely data-driven M0 model performs well in the early stages of training data coverage (…),… t The prediction curve (<30h) can fit the experimental points well, but in the subsequent data-free prediction interval, its prediction curve is severely distorted (e.g., Figure 9 (The solid green line in the middle) This phenomenon reveals that when the "black box" lacks physical constraints, it cannot learn the intrinsic evolution law of rock deformation over time by simply memorizing the distribution of training samples, and therefore does not have the ability to extrapolate time.
[0076] Conversely, the prediction curve of the HC-PINN model ( Figure 9 The blue dashed line (in the middle) matches the theoretical solution curve throughout the entire loading period (0–303.6 h). Figure 9 (in red solid line) and discrete test data points ( Figure 9 The black circles in the diagram show extremely high consistency. This indicates that the hard-constraint architecture and robust optimization strategy successfully internalize the generalized Kelvin constitutive relation into the prior knowledge of the neural network, enabling the model to still reason according to physical laws in areas without data.
[0077] This invention also provides a hard-constrained physical information neural network prediction system for time-dependent deformation of rock mass, applicable to the above-mentioned method, including: The neural network building module is used to build deep neural networks that take normalized spatiotemporal coordinates as input and displacement vectors as output. The hard constraint module is used to multiply each component of the displacement vector by the corresponding geometric distance factor to obtain the predicted displacement field. The geometric distance factor is zero on the target boundary, so that the predicted displacement field satisfies the Dirichlet boundary condition on the target boundary. The automatic differential physics engine module is used to calculate the partial derivatives of the predicted displacement field and substitute them into the rock rheological constitutive equation to calculate the physical residuals; The loss function construction module is used to construct the total loss function by weighted summation of the deviation between the predicted displacement field and the monitoring data and the physical residuals using a robust loss function. The adaptive optimization module is used to minimize the total loss function to train the deep neural network, and after convergence, the predicted displacement field is used as the prediction result.
[0078] The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described; only preferred embodiments of the present invention are illustrated. The descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of the present invention. As long as the combination of these technical features does not contradict each other, it should be considered within the scope of this specification.
[0079] It should be noted that those skilled in the art can make various modifications and improvements without departing from the inventive concept, and these all fall within the scope of protection of this invention. Therefore, the scope of protection of this invention should be determined by the appended claims.
Claims
1. A hard-constrained physical information neural network prediction method for time-dependent deformation of rock mass, characterized in that, Includes the following steps: Step S1: Construct a deep neural network with normalized spatiotemporal coordinates as input and displacement vector as output; Step S2: Multiply each component of the displacement vector by the corresponding geometric distance factor to obtain the predicted displacement field. The geometric distance factor is zero on the target boundary, so that the predicted displacement field satisfies the Dirichlet boundary condition on the target boundary. The geometric distance factor is defined in the solution domain. Generalized distance function on The generalized distance function Satisfy: When the spatial coordinates point Located in the solution domain boundary When, When spatial coordinates point Located in the solution domain Inside, ; For the solution domain of a cylindrical specimen under axisymmetric conditions, the target boundary includes the axis of symmetry and the bottom end face. The generalized distance function is taken as the ratio of the radial coordinate to the specimen width and the ratio of the axial coordinate to the specimen height, respectively. The radial predicted displacement of the predicted displacement field... and axial predicted displacement They are defined as follows: ; ; in, Radial coordinates, For axial coordinates, For time, The width of the sample. The height of the sample. and These are the radial and axial displacement components output by the deep neural network, respectively. This is the set of trainable parameters for a deep neural network. Step S3: Calculate the partial derivatives of the predicted displacement field and substitute them into the rock rheological constitutive equation to calculate the physical residuals; The rheological constitutive equation of the rock is the three-dimensional deviatoric stress-deviatoric strain tensor differential equation of the generalized Kelvin model, expressed as: ; in, These are components of the deviatoric stress tensor. These are the components of the deviatoric strain tensor. The elastic shear modulus of the Hooke body. The elastic shear modulus of the Kelvin body. The shear viscosity coefficient of Kelvin volume. express directional surface along The deviatoric stress component acting in the direction of the stress. express direction and Deviatoric strain components between directions, subscript and Values and ,in Radial, For the axial direction; The physical residuals include constitutive equation residuals. Stress boundary residuals Poisson effect residuals Three components; The constitutive equation residual The expression is: ; The stress boundary residual The expression is: ; The Poisson effect residual The expression is: ; In the formula, Predict stress at the boundary. For the true boundary stress, The radial strain value is calculated based on the predicted displacement field. The axial strain value is calculated based on the predicted displacement field. Poisson's ratio; Step S4: The deviation between the predicted displacement field and the monitoring data, as well as the physical residual, are weighted and summed using a robust loss function to construct a total loss function; Step S5: Minimize the total loss function to train the deep neural network, and use the predicted displacement field as the prediction result after convergence.
2. The method according to claim 1, characterized in that, The total loss function described in step S4 Represented as: ; in, This is the set of trainable parameters for a deep neural network. The data loss term is calculated using the Huber loss function based on the deviation between the predicted displacement field and the monitoring data. The physical loss term is calculated using the Smooth-L1 loss function on the physical residuals. and These are the weighting coefficients for the data loss term and the physical loss term, respectively.
3. The method according to claim 2, characterized in that, The weighting coefficient and Adaptive adjustment is performed using a gradient norm-based dynamic adjustment algorithm: the data loss term is recorded in real time during training. and the physical loss term Each historical decline rate is used to adaptively adjust the weight coefficient of the physical loss term based on the ratio of the historical decline rates at preset fixed iteration intervals. This allows the data loss term and the physical loss term to converge collaboratively throughout the training process.
4. The method according to claim 1, characterized in that, The collocation points used in step S3 to calculate the physical residual are generated using a spatiotemporal non-uniform sampling strategy. Specifically, the spatial domain collocation points are generated on the sample cross section using Latin hypercube sampling technology, and the time domain collocation points are generated using a mixed logarithmic distribution sampling strategy.
5. The method according to claim 1, characterized in that, The deep neural network includes an input layer, multiple fully connected hidden layers, and an output layer; the input layer receives three-dimensional normalized spatiotemporal coordinates composed of radial coordinates, axial coordinates, and time; the neuron output of each fully connected hidden layer undergoes a nonlinear transformation through a hyperbolic tangent activation function, and the output of the neuron in the first layer is... Layer output vector Represented as: ; in, and The first Layer weight matrix and bias vector, The activation function is hyperbolic tangent; the weight parameters of the deep neural network are initialized using the Xavier normal distribution strategy. Step S5, when minimizing the total loss function, also includes the following training strategy: Gradient clipping: Before each parameter update, the gradient norm of the total loss function with respect to the deep neural network is calculated, and the gradient is clipped when the gradient norm exceeds a preset clipping threshold. Dynamic learning rate scheduling: A learning rate scheduling strategy combining a linear warm-up phase and a cosine annealing phase is adopted. In the linear warm-up phase, the learning rate increases linearly from zero to the peak learning rate. In the cosine annealing phase, the learning rate decays from the peak learning rate to the minimum learning rate according to the cosine function curve.
6. A hard-constrained physical information neural network prediction system for time-dependent deformation of rock mass, applicable to the method described in any one of claims 1 to 5, characterized in that, include: The neural network building module is used to build deep neural networks that take normalized spatiotemporal coordinates as input and displacement vectors as output. The hard constraint module is used to multiply each component of the displacement vector by the corresponding geometric distance factor to obtain the predicted displacement field. The geometric distance factor is zero on the target boundary, so that the predicted displacement field satisfies the Dirichlet boundary condition on the target boundary. An automatic differential physics engine module is used to calculate the partial derivatives of the predicted displacement field and substitute them into the rock rheological constitutive equation to calculate the physical residuals; The loss function construction module is used to construct a total loss function by weighted summation of the deviation between the predicted displacement field and the monitoring data and the physical residual using a robust loss function. An adaptive optimization module is used to minimize the total loss function to train the deep neural network, and after convergence, the predicted displacement field is used as the prediction result.
Citation Information
Patent Citations
Urban block wind environment rapid prediction method by embedding physical information into generative adversarial network
CN120257798A
Four-side gridding-oriented multi-target distance field calculation method based on physical information neural time field
CN120494238A
