High-precision and high-efficiency calculation method and system for laser welding numerical model based on physical information neural network

Abstract

The invention belongs to the field of laser welding numerical simulation, and discloses a laser welding numerical model high-precision efficient calculation method and system based on a physical information neural network. Generating model basic training data based on the laser welding multiphase flow numerical model; a 3D ConvLSTM-based laser welding physical field prediction model is constructed, energy input information is introduced into a hidden layer, and the randomness of the laser welding process is restored; designing a loss function containing various error sources; adjusting the network weight by using a gradient minimization algorithm, and minimizing a loss function; and inputting the three-dimensional physical field data of T moments to predict the change of subsequent T moments, extracting the data of (T-T) moments after the input data and the output data to reform a model for input, and predicting the physical field change of the subsequent moments. The invention provides a high-precision and high-efficiency novel solving method for laser welding numerical model calculation.

Description

Technical Field

[0001] The present invention belongs to but is not limited to the field of laser processing numerical simulation technology, and in particular relates to a high-precision and high-efficiency calculation method and system for a laser welding numerical model based on a physical information neural network. Background Art

[0002] The laser welding process involves a variety of complex physical phenomena, including the interaction between the laser and the metal, the melting and solidification of the metal, and heat transfer. It is characterized by strong time-varying, nonlinear, and high-dimensional characteristics. It is a typical multi-scale, multi-physics field coupled gas-liquid-solid multiphase flow evolution process. Numerical simulation technology, by constructing a mathematical model to reconstruct the laser welding process, enables a visual description of the welding process. It is an important technical means to study the evolution of the laser welding molten pool and keyhole and optimize process parameters.

[0003] Solving numerical models for laser welding can be simply summarized as solving a system of partial differential equations (PDEs). Traditional numerical methods, such as the finite element method (FEM), finite difference method (FDM), and finite volume method (FVM), decompose the computational domain into several discrete grids and iteratively solve the discrete data points on the grid to approximate the solution to the system of equations. However, to ensure the accuracy and reliability of the numerical analysis, more precise meshing and tighter time steps must be used. This results in a significant consumption of computational time and resources for single numerical simulations, a major obstacle to the practical application of laser welding numerical simulation technology. One current mainstream approach is to use parallel computing to address the long computational time. While this approach yields accurate numerical results, it still incurs a significant computational resource overhead. Another approach is to use neural network technology to construct predictive models to efficiently and quickly obtain the desired results. However, these methods operate as black boxes and lack physical insights. Furthermore, the construction of neural network models requires a large amount of data and places stringent demands on data quality. High-quality, large-sample data is crucial for maintaining high prediction accuracy.

[0004] In view of the above analysis, the technical problems that need to be solved urgently in the existing technology are:

[0005] How to simultaneously take into account the high precision of numerical methods and the high efficiency of neural network models, use fewer samples to construct a new computational solution method with clear physical and guiding significance, and provide a "high precision" and "efficient" laser welding numerical model calculation method for the field of laser welding numerical simulation has become an urgent problem to be solved. Summary of the Invention

[0006] In response to the problems existing in the prior art, the present invention provides a high-precision and high-efficiency calculation method and system for a laser welding numerical model based on a physical information neural network.

[0007] The present invention is achieved by providing a high-precision and high-efficiency calculation method for a laser welding numerical model based on a physical information neural network, characterized in that the high-precision and high-efficiency calculation method for a laser welding numerical model based on a physical information neural network specifically includes:

[0008] S1: Generate model basic training data based on the laser welding multiphase flow numerical model;

[0009] S2: Build a 3D ConvLSTM-based laser welding physical field prediction model, introduce energy input information in the hidden layer, and restore the randomness of the laser welding process;

[0010] S3: Design a loss function that includes multiple error sources, including physical information error terms (physical control equation residuals), data error terms, boundary condition error terms, and initial condition error terms;

[0011] S4: Output the prediction results, use the gradient minimization algorithm to adjust the network weights, calculate the loss function, and minimize the loss function;

[0012] S5: Input the three-dimensional physical field data at T moments to predict the changes at the subsequent T' moments, extract the data at (TT') moments after the input data and recompose it with the output data to form the model input, and predict the physical field changes at the subsequent moments.

[0013] Furthermore, the basic training data of the model described in S1 is obtained, and the low-precision laser welding multiphase flow numerical model is constructed by modeling various physical phenomena in the laser welding process, including solid-liquid-gas three-phase heat transfer, flow behavior (including solid-liquid phase change), multiple reflection heat source model, etc., and solving the model using fluid calculation software; extracting the physical field time series data in the calculation results, including temperature field, flow field, and fluid pressure field under different welding processes and material parameters;

[0014] The model input data format is: [sample batch size, input time step, X-axis coordinate, Y-axis coordinate, Z-axis coordinate, number of data channels], and [B, T, X, Y, Z, C], where B represents the sample batch size, T represents the input time step, X, Y, Z represent the three-dimensional spatial coordinates of the data points respectively, and C represents the number of data channels. The value of C in the present invention is 5, including temperature, X, Y, Z direction velocity components, and fluid pressure; the model output data format is: [sample batch size, output time step, X-axis coordinate, Y-axis coordinate, Z-axis coordinate, number of data channels], and [B, T`, X, Y, Z, C], where T` represents the output time step; the model of the present invention inputs physical field time series data of length T and outputs physical field time series prediction results of length T`; the model's process parameters such as laser power, welding speed, defocus amount, etc., and the model's material properties such as density of each phase material, specific heat capacity, thermal conductivity, solid-liquid phase line temperature, liquid metal viscosity, etc. are input into the model together as macroscopic invariants.

[0015] Furthermore, the step S2 specifically includes the following steps:

[0016] S2.1: Adjust the training data format to the input format and input T-step 3D physical time series data into the model;

[0017] S2.2: The model uses multiple layers of 3D ConvLSTM layers to simultaneously extract the spatiotemporal features of 3D physical time series data using 3D convolution kernels to achieve time series prediction of 3D physical data for laser welding.

[0018] S2.3: Introduce energy input information into the 3D ConvLSTM layer and calculate and extract the fluid volume fraction to obtain the gas-liquid interface distribution. The fluid volume fraction calculation formula is as follows:

[0019]

[0020] in the formula Input the flow field data of the model at the current moment, and use the gas-liquid interface distribution combined with information such as laser power and beam shape to calculate the energy input information of the current time step using the multiple reflection heat source model. The formula is as follows:

[0021] The energy absorption formula of a certain reflection point is as follows:

[0022] Q=I0(r)*α F

[0023] Where I0(r) is the intensity of the laser beam at the reflection point, in W / mm2, α F is the laser absorption coefficient. The light intensity distribution calculation formula is as follows:

[0024]

[0025] Where P is the laser power, is the radius constant of the laser beam focal plane, and r is the distance between the reflection point and the origin of the laser beam plane. The laser absorption coefficient is calculated as follows:

[0026]

[0027] Where ε is the Fresnel absorption coefficient and φ is the angle between the incident light beam and the normal vector of the reflection plane. The calculation of the normal vector of the reflection plane is based on the fluid volume fraction and the calculation formula is as follows:

[0028]

[0029] The formula for the beam reflection direction vector is as follows:

[0030]

[0031] in, is the beam reflection direction vector, N i is the direction vector of the incident beam, n is the normal vector of the reflection interface, and the model assumes that the laser beam will no longer reflect when the energy is less than 10%;

[0032] S2.4: After introducing the energy input information, add a 3D ConvLSTM layer to generate the time-recursive prediction results of the laser welding physical field based on the spatiotemporal distribution characteristics and energy input information; finally, add a 3D convolution layer to map the hidden channels back to the five physical quantity channels.

[0033] Furthermore, the loss function designed in S3 includes multiple error sources, and the physical information error includes: mass conservation equation, momentum conservation equation and energy conservation equation. The prediction result is substituted into the above partial differential equation to calculate the residual to form a part of the loss function, which is used to constrain the prediction result to conform to physical laws and improve the physical consistency of the prediction result; the data error term is the difference between the prediction result and the standard value of the training set, and the calculation of the boundary condition and initial condition error is the same as above; the model parameters are adjusted according to the loss function results, including error weight, input and output time steps, learning rate and other hyperparameters.

[0034] Furthermore, the laser welding physical field prediction model training process in S4 specifically includes the following steps:

[0035] S4.1: Input three-dimensional physical field data (velocity, pressure, temperature) and related process and material properties at T time points;

[0036] S4.2: Predict the physical field data for the next T' time periods using the energy input information addition method described in S2 and the laser welding physical field prediction model;

[0037] S4.3: Use gradient minimization algorithm to adjust network weights;

[0038] S4.4: Calculate the loss function and minimize the loss function.

[0039] Furthermore, the S5, full-seam laser welding physical field evolution behavior prediction process, specifically includes the following steps:

[0040] S5.1: Input the three-dimensional physical field data and related process and material property parameters at T moments under different process parameters and material property conditions;

[0041] S5.2: Use the steps described in S2 to predict the evolution of the laser welding physical field at the next T moments;

[0042] S5.3: Extract the data (TT`) after the input data and the output data to reconstruct the model input and predict the physical field changes at subsequent times;

[0043] S5.4: Through the above process, the efficient solution of the laser welding numerical model is achieved, and the accurate prediction of the physical field evolution behavior during the full weld laser welding process is restored.

[0044] Another object of the present invention is to provide a high-precision and high-efficiency calculation system for a laser welding numerical model based on a physical information neural network, the system specifically comprising:

[0045] A data preparation module is used to generate a certain amount of laser welding physical data that meets the physical laws using a low-precision laser welding multiphase flow numerical model;

[0046] Model building module, used to build a laser welding physical field prediction model based on 3D ConvLSTM;

[0047] Training module, used to train the laser welding physical field prediction model;

[0048] Prediction module, used to predict the evolution behavior of the physical field in full-seam laser welding.

[0049] In combination with the above technical solutions and the technical problems solved, the advantages and positive effects of the technical solutions to be protected by the present invention are as follows:

[0050] The fluctuating behavior of laser welding keyholes is highly random. This invention accurately restores this randomness by introducing a multiple reflection model into the neural network model to calculate energy input information. This adds a practical physical insight to the model training process, standardizes the prediction results, improves model training efficiency, and lays the foundation for achieving high-precision prediction with a small number of samples.

[0051] This paper trains the model by introducing three types of biases: observation bias introduced by the model's basic training data, inductive bias introduced by energy information input, and learning bias introduced by the loss function due to multiple error sources. This achieves high-precision predictions with limited sample data. The additional guidance provided by these multiple biases is of great significance for training neural network models in situations where data is scarce or expensive to acquire.

[0052] The present invention improves the generalization ability of the model by adding physical insights to the data-driven model, solves the problem of large prediction deviation caused by data differences in the extrapolation of previous data-driven models, and provides a "high-precision" and "efficient" prediction method for the laser welding physical field driven by data-mechanism coupling.

[0053] This invention fills a certain technical gap in the field of efficient calculation in the field of laser welding numerical simulation. Specifically, it provides a high-efficiency numerical model calculation method for improving the calculation efficiency of the numerical model.

[0054] As the numerical model of laser welding becomes more complex, the computational difficulty and consumption of computing resources continue to increase. At the same time, machine learning technology is an efficient means of physical field prediction, and researchers have been trying to introduce it into the calculation of numerical models to improve computational efficiency. However, the high cost of data acquisition and large prediction deviations have prevented the above ideas from being successful. The technical solution of the present invention provides a novel method for solving the above technical problems, namely, by introducing three different types of deviations into the neural network model, efficient model training and accurate prediction under few sample conditions are achieved.

[0055] The technical solution of the present invention overcomes the biases of the two technical approaches, realizing a computational method for numerical simulation of laser welding that combines both high efficiency and high precision. Specifically, it overcomes the low efficiency of traditional numerical simulation methods for laser welding by introducing machine learning into numerical calculations and utilizing a physical information neural network model to achieve efficient computation and prediction of the laser welding physical field. It also overcomes the high sample requirements and large prediction deviations of traditional neural network models by introducing three types of deviations during the training process of the neural network model, achieving high-precision predictions in scenarios with few samples or high data acquisition costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 This is a flow chart of a high-precision and high-efficiency calculation method for a laser welding numerical model based on a physical information neural network provided by an embodiment of the present invention;

[0057] Figure 2 This is a module diagram of a high-precision and high-efficiency calculation system for a laser welding numerical model based on a physical information neural network provided by an embodiment of the present invention;

[0058] Figure 3 Schematic diagram of a low-precision laser welding multiphase flow numerical model provided by an embodiment of the present invention;

[0059] Figure 4 This is the architecture and training flow chart of the laser welding physical field prediction model provided by an embodiment of the present invention;

[0060] Figure 5 This is a flow chart of the prediction of the physical field evolution behavior of full-weld laser welding provided by an embodiment of the present invention;

[0061] Figure 6 This is a three-dimensional prediction result diagram of a certain example provided by an embodiment of the present invention;

[0062] Figure 7 This is a diagram of the time series prediction results of a two-dimensional molten pool-keyhole cross section of a certain example provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0063] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0064] Laser welding numerical simulations often rely on high-precision multiphase flow solvers to obtain information about the melt pool temperature field, flow field, and phase change. However, these solvers are computationally intensive and converge slowly, making them difficult to meet the dual requirements of "high precision and high efficiency" in industrial parameter optimization, real-time monitoring, and digital twin platforms. Furthermore, traditional data-driven alternative models often lack self-consistent constraints on physical laws and are prone to drift under boundary conditions and extreme operating conditions. This makes it impossible to guarantee the reliability and interpretability of predictions, limiting their practical application on production lines.

[0065] To address these bottlenecks, we first utilized a mature numerical model of multiphase flow in laser welding to perform parametric sweeps of typical process parameters (such as laser power, scanning speed, and focal spot size), generating high-resolution spatiotemporal data covering various welding scenarios. By normalizing the simulation results and dividing them into training, validation, and test sets, we constructed a 3D volumetric dataset that simultaneously reflects temperature gradients, melt pool streamlines, and phase transformation trends, providing sufficient and diverse physical samples for subsequent network learning.

[0066] In terms of network design, a three-dimensional convolutional long-short-term memory (LSTM) architecture is introduced to simultaneously capture spatial neighborhood characteristics and temporal evolution patterns. The network feeds the physical field volume data from the previous few frames into a multi-layer 3D-Conv unit to extract local thermal-fluid coupling features. LSTM units then implement temporal memory in the hidden layer. The laser energy distribution encoding information is then injected into the gating mechanism, ensuring that the model can perceive the differences in melt pool reactions under different energy inputs and retain a memory of the process dynamics for long-term predictions.

[0067] To balance physical consistency with data fitting accuracy, the present invention incorporates a composite loss function that combines the residuals of the physical governing equations, simulation data errors, boundary condition errors, and initial condition errors. The residuals of the energy conservation and momentum equations on the prediction field are calculated through automatic differentiation and minimized together with the mean square error of the label field. Additional boundary and initial time error terms are introduced to achieve strict constraints on the boundary heat flux and initial temperature distribution, significantly reducing the risk of overfitting without physical support.

[0068] During model training, gradient descent optimization is performed using small batches of time series data. Learning rate warm-up and cosine annealing are combined to improve convergence stability, and gradient clipping is used to prevent gradient explosion. To address gradient decay during backpropagation over long sequences, the truncated backpropagation through time (BPTT) technique is used to segment long-term memory, effectively modeling multi-step evolution within limited video memory. This allows convergence to a satisfactory physical error level within dozens of iterations.

[0069] During deployment, the model takes historical physical field data and laser parameter sequences as input, gradually outputs 3D field predictions for the next moment, and then rolls these predictions into the input buffer, enabling rapid time-series deduction. Compared to traditional numerical solutions, this method reduces computation time by dozens of times while maintaining submillimeter spatial resolution and sub-second timestep accuracy. This provides a viable, high-precision and efficient solution for online quality monitoring, real-time process parameter correction, and digital twin systems.

[0070] like Figure 1 As shown, an embodiment of the present invention provides a high-precision and high-efficiency calculation method for a laser welding numerical model based on a physical information neural network, the method specifically comprising:

[0071] S1: Acquisition of basic model training data. Based on the established low-precision laser welding multiphase flow numerical model, high-quality laser welding physical data that meets physical laws is generated, including temperature field, flow field, and fluid pressure field under different welding processes and material parameters.

[0072] S2: Build a laser welding physical field prediction model based on 3D ConvLSTM. The model uses the input T-step 3D physical time series data to predict the changes in the physical field in the subsequent T' steps.

[0073] S3: Design a loss function that includes multiple error sources, including physical information error terms (physical control equation residuals), data error terms, boundary condition error terms, and initial condition error terms;

[0074] S4: training the laser welding physics prediction model;

[0075] S5: Predicting the physical field evolution behavior of full-seam laser welding.

[0076] The specific explanation of the acquisition of the model basic training data in S1 is as follows:

[0077] The low-precision laser welding multiphase flow numerical model is constructed by modeling various physical phenomena in the laser welding process, including solid-liquid-gas three-phase heat transfer, flow behavior (including solid-liquid phase change), and multiple reflection heat source models. The model is solved using fluid calculation software. The physical field time series data in the calculation results are extracted, including the temperature field, flow field, and fluid pressure field under different welding processes and material parameters.

[0078] The purpose of solving low-precision numerical models is to quickly obtain a batch of basic training data for the model that meets the physical laws. This method has relatively little dependence on data quality and data volume, but still requires a certain amount of training data to complete the model training work;

[0079] The model input data format is: [sample batch size, input time step, X-axis coordinate, Y-axis coordinate, Z-axis coordinate, number of data channels], and [B, T, X, Y, Z, C], where B represents the sample batch size, T represents the input time step, X, Y, Z represent the three-dimensional spatial coordinates of the data point, C represents the number of data channels, and the value of C in the present invention is 5, including temperature, X, Y, Z direction velocity components, and fluid pressure; the model output data format is: [sample batch size, output time step, X-axis coordinate, Y-axis coordinate, Z-axis coordinate, number of data channels], and [B, T', X, Y, Z, C], where T' represents the output time step; the model of the present invention inputs physical field time series data of length T and outputs physical field time series prediction results of length T'; in order to effectively train the model, the model's process parameters such as laser power, welding speed, defocus, etc., and the model's material properties such as density of each phase material, specific heat capacity, thermal conductivity, solid-liquid phase line temperature, liquid metal viscosity, etc. are input into the model together as macroscopic invariants;

[0080] In S2, a 3D convolution kernel is used in each ConvLSTM layer to extract the three-dimensional spatiotemporal features of the physical space in the time series data. Time series prediction of physical information is achieved through multiple 3D ConvLSTM layers. Energy input information is introduced into a hidden layer, and the gas-liquid interface distribution is obtained by calculating and extracting the fluid volume fraction. The energy input information for the current time step is calculated using a multi-reflection heat source model, combining information such as laser power and beam shape. The keyhole's wave behavior is affected by the energy input. By introducing this physical information, a randomness is introduced into the physical field prediction to improve the model's generalization ability. Specifically, the following steps are included:

[0081] S2.1: Adjust the training data format to the input format and input T-step 3D physical time series data into the model;

[0082] S2.2: The model uses multiple layers of 3D ConvLSTM layers to simultaneously extract the spatiotemporal features of 3D physical time series data using 3D convolution kernels to achieve time series prediction of 3D physical data for laser welding.

[0083] S2.3: Introduce energy input information into the 3D ConvLSTM layer and calculate and extract the fluid volume fraction to obtain the gas-liquid interface distribution. The fluid volume fraction calculation formula is as follows:

[0084]

[0085] in the formula Input the flow field data of the model at the current moment. Utilize the gas-liquid interface distribution combined with information such as laser power and beam shape to calculate the energy input information of the current time step using the multi-reflection heat source model. The formula is as follows:

[0086] The energy absorption formula of a certain reflection point is as follows:

[0087] Q=I0(r)*α F

[0088] Where I0(r) is the intensity of the laser beam at the reflection point, in W / mm2, α F is the laser absorption coefficient. The light intensity distribution calculation formula is as follows:

[0089]

[0090] Where P is the laser power, is the radius constant of the laser beam focal plane, and r is the distance between the reflection point and the origin of the laser beam plane. The laser absorption coefficient is calculated as follows:

[0091]

[0092] Where ε is the Fresnel absorption coefficient and φ is the angle between the incident light beam and the normal vector of the reflection plane. The calculation of the normal vector of the reflection plane is based on the fluid volume fraction and the calculation formula is as follows:

[0093]

[0094] The formula for the beam reflection direction vector is as follows:

[0095]

[0096] in, is the beam reflection direction vector, N i is the direction vector of the incident beam, and n is the normal vector of the reflection interface. The model assumes that the laser beam will no longer reflect when the energy is less than 10%. This heat source model accurately reflects the actual motion trajectory of the laser beam during laser welding, providing a basis for accurate energy distribution calculations.

[0097] S2.4: After introducing the energy input information, a 3D ConvLSTM layer is added to generate a time-recursive prediction result of the laser welding physical field based on the spatiotemporal distribution characteristics and the energy input information. Finally, a 3D convolutional layer is added to map the hidden channel back to the five physical quantity channels.

[0098] The loss function designed in S3 contains multiple error sources and specifically includes the following parts:

[0099] Physical information errors include: mass conservation equations, momentum conservation equations, and energy conservation equations. The prediction results are substituted into the above partial differential equations to calculate the residuals, which constitute part of the loss function. This is used to constrain the prediction results to conform to physical laws, thereby improving the physical consistency of the prediction results. The data error term is the difference between the prediction results and the standard values ​​of the training set. The calculation of boundary condition and initial condition errors is the same as above. The model parameters are adjusted according to the loss function results, including hyperparameters such as error weight, input and output time steps, and learning rate.

[0100] Physical information error term (physical control equation residual):

[0101] Physical information errors include: mass conservation equation, momentum conservation equation, and energy conservation equation. The specific calculation methods for the three error sources are as follows:

[0102] Residual of the mass conservation equation:

[0103]

[0104] in, It represents the velocity component value of the grid with coordinates (i, j, k) at the nth moment. Δx, Δy, and Δz are increments in three directions, which are determined by the grid size.

[0105] Residual of the momentum conservation equation:

[0106]

[0107] In order to simplify the formula, S refers to the velocity components in different directions (1, 2, 3 refer to the velocity components u, v, w in order). Indicates the pressure value of the grid with coordinates (i, j, k) at the nth moment. Indicates the forces involved in the laser welding process, including surface tension, thermal buoyancy, recoil pressure, etc.

[0108] The surface tension calculation formula is as follows:

[0109] P σ =κσ

[0110] Where κ is the curvature of the free surface and σ is the surface tension coefficient;

[0111] The recoil pressure calculation formula is as follows:

[0112]

[0113] Where P0 is the atmospheric pressure, L v is the latent heat of vaporization, T b is the vaporization temperature, is the ideal gas constant, Indicates the temperature value of the grid with coordinates (i, j, k) at the nth moment;

[0114] The calculation formula for thermal buoyancy is as follows:

[0115]

[0116] Where β is the thermal expansion coefficient, g is the acceleration due to gravity, and T l is the melting point of the material;

[0117] Residual of the energy conservation equation:

[0118]

[0119] Where k is thermal conductivity, c p is the specific heat capacity, S is the energy source term obtained by the multiple reflection heat source model, It represents the thermal enthalpy of the grid with coordinates (i, j, k) at the nth moment. The calculation formula is as follows:

[0120]

[0121] where ρ s is the solid phase density of the material, ρ l is the liquid phase density of the material, cs and c l are the specific heat of solid and liquid phases per unit volume, T s and T l are the solidus and liquidus of the material, ΔH l is the latent heat of fusion for solid phase change.

[0122] Data error term:

[0123]

[0124] The data error term is the difference between the model prediction data and the actual model numerical solution data;

[0125] Boundary condition error term:

[0126] Radiation heat transfer boundary condition error:

[0127]

[0128] where h A is the convective heat transfer coefficient, T ∞ is the ambient temperature, generally set to 293K, σ s is the Stefan-Boltzmann constant, ε r is the emissivity, Q vap is the vaporization heat loss, (a, b, c) is the normal vector of the boundary plane;

[0129] Mechanical boundary condition error:

[0130]

[0131] where γ and R represent the surface tension coefficient and the surface curvature radius.

[0132] Initial condition error:

[0133]

[0134] The total loss function formula is as follows:

[0135]

[0136] The laser welding physical field prediction model training process in S4 specifically includes the following steps:

[0137] S4.1: Input three-dimensional physical field data (velocity, pressure, temperature) and related process and material properties at T time points;

[0138] S4.2: Predict the physical field data for the next T' time periods using the energy input information addition method described in S2 and the laser welding physical field prediction model;

[0139] S4.3: Use gradient minimization algorithm to adjust network weights;

[0140] S4.4: Calculate the loss function and minimize the loss function;

[0141] The S5, prediction process of the physical field evolution behavior of full-seam laser welding, specifically includes the following steps:

[0142] S5.1: Input the three-dimensional physical field data and related process and material property parameters at T moments under different process parameters and material property conditions;

[0143] S5.2: Use the steps described in S2 to predict the evolution of the laser welding physical field at the next T moments;

[0144] S5.3: Extract the data (TT`) after the input data and the output data to reconstruct the model input and predict the physical field changes at subsequent times;

[0145] S5.4: Through the above process, the efficient solution of the laser welding numerical model is achieved, and the accurate prediction of the physical field evolution behavior during the full weld laser welding process is restored.

[0146] like Figure 2 As shown, an embodiment of the present invention provides a high-precision and high-efficiency calculation system for a laser welding numerical model based on a physical information neural network, specifically comprising:

[0147] A data preparation module is used to generate a certain amount of laser welding physical data that meets the physical laws using a low-precision laser welding multiphase flow numerical model;

[0148] Model building module, used to build a laser welding physical field prediction model based on 3D ConvLSTM;

[0149] Training module, used to train the laser welding physical field prediction model;

[0150] Prediction module, used to predict the evolution behavior of the physical field in full-seam laser welding.

[0151] Embodiments of the present invention provide a high-precision and efficient calculation method for a laser welding numerical model based on a physical information neural network. Based on physical information neural network (PINN) technology, high-precision and efficient calculation of the laser welding numerical model is achieved by adding physical information of various deviation types at multiple stages of model construction. In the data preparation stage, the present invention utilizes a low-precision laser welding multiphase flow numerical model to generate a certain amount of laser welding physical data that conforms to physical laws, including temperature fields, flow fields, and fluid pressure fields under different welding parameters, thereby introducing observation bias into the calculation model. In the model definition stage, the present invention introduces laser energy input information into a hidden layer. This physical information considers the energy input differences caused by different welding processes and accurately calculates the welding keyhole energy distribution by extracting the fluid volume fraction distribution at the current time step using a multiple reflection heat source model, thereby introducing inductive bias into the calculation model. In the calculation loss stage, the present invention designs a loss function that incorporates multiple error sources, including: physical information error terms (physical control equation residuals), data error terms, boundary condition error terms, initial condition error terms, etc., thereby introducing learning bias into the calculation model. The above model is verified using numerical simulation results under other welding parameters. By introducing physical information of different deviation types into the deep neural network model, high-precision numerical model solution is achieved under few-sample conditions. Combined with the 3D ConvLSTM neural network model, time series prediction of three-dimensional physical data is realized, providing a new high-precision and efficient solution method for laser welding numerical model calculation.

[0152] The data involved in this embodiment comes from a low-precision laser welding multiphase flow numerical model diagram as shown in the following figure: Figure 3 As shown. The commercial software FLOW 3D is used to solve the numerical model. A set of three-dimensional physical field data of laser welding with an initial time of 1μs, a time step of 0.01μs, and a total of 4μs is obtained, and the total number of three-dimensional calculation domains reaches 50,000. The input data format is adjusted to: [B, T, X, Y, Z, C], B is the sample batch size, 8 is used in this embodiment, and T is the input time step, 12 is used in this case. The output data format of this embodiment is: [B, T`, X, Y, Z, C], T` is the output time step, and 6 is used in this embodiment.

[0153] The architecture and training flow chart of the laser welding physical field prediction model involved in this embodiment are as follows: Figure 4 As shown. Input the three-dimensional physical field data (velocity, pressure, temperature) and related process and material property parameters at T time points. This example uses 2024-T3 aluminum alloy as the research material. The relevant material parameters serve as macroscopic invariants to supplement the physical information of the model. The specific values ​​are shown in Table 1.

[0154] Table 1 Thermophysical properties of 2024-T3 aluminum alloy

[0155]

[0156] This embodiment uses three layers of 3D ConvLSTM layers to simultaneously extract the spatiotemporal features of three-dimensional physical time series data using three-dimensional convolution kernels. The velocity data of the input data is extracted to calculate the fluid volume fraction, and the energy input information is calculated based on the multiple reflection model and input into the model as a feature. Then, a layer of 3DConvLSTM layer is used to generate the time recursive prediction results of the laser welding physical field based on the spatiotemporal distribution characteristics and energy input information. Finally, a layer of 3D convolution layer is added to map the hidden channel back to the five physical quantity channels. The gradient minimization algorithm is used to adjust the network weights. The loss function is calculated according to the loss function calculation method in S3, and the loss function is minimized.

[0157] The prediction process of the physical field evolution behavior of full weld laser welding in this embodiment is as follows: Figure 5 As shown in the figure, the 3D physical field data and related process and material property parameters at T time points under different process parameters and material property conditions are input. The model is used to predict the physical field evolution at the next T' time points. The data (TT') after the input data is extracted and reassembled with the output data to form the model input, and the physical field changes at subsequent time points are predicted. This process enables the efficient solution of the laser welding numerical model and the accurate prediction of the physical field evolution behavior during full-seam laser welding.

[0158] Using the above method, we get a case, and the prediction results are compared with Figure 6 、 7 shown.

[0159] The present invention provides a high-precision and high-efficiency method for the calculation of laser welding numerical models. Specifically, it is embodied in: using the physical information neural network model, introducing three types of deviations, and realizing the prediction of the laser welding physical field under the conditions of few samples or high data acquisition costs. Different from the previous numerical solution method, the present invention is an efficient calculation method based on machine learning. In actual application scenarios such as: titanium alloy laser welding formation prediction, this method uses a small amount of simulation data to train the prediction model, realizes the efficient prediction of the entire weld formation, greatly shortens the research cycle, and overcomes the limitations of traditional methods. Specific effects such as Figure 6 、 7 At the same time, the present invention has a certain degree of flexibility. For different welding process parameters, the model prediction results can be adjusted by adjusting the energy information input part.

[0160] The prediction results of the laser welding physical field evolution described in the present invention are as follows Figure 6 、 7 The technical effect obtained by using the present invention is also shown. Figure 6The keyhole evolution and molten pool distribution during laser welding can be observed. To highlight this effect, the present invention shows the two-dimensional molten pool-keyhole cross-section time series prediction results as follows: Figure 7 As shown, the evolution of the keyhole and the morphology of the molten pool during laser welding are also shown. The above prediction results are also recorded in other literatures, proving the feasibility of the scheme of the present invention. From a theoretical analysis, the loss function of the various error sources described in the present invention includes four parts: physical information error term, data error term, boundary condition error term, initial condition error term, etc. Among them, the physical information error term is based on the three major physical conservation equations, and the equation is discretized using the finite difference method. The boundary conditions take into account the influence of radiation heat transfer and surface tension. The above loss function accurately restores the physical phenomena in the laser welding process and can effectively constrain the prediction results of the neural network model so that it meets the physical laws. The above equations are also recorded in other literatures on numerical simulation of laser welding. The energy information input described in the present invention uses a multiple reflection heat source model to accurately restore the heat input in the laser welding process, providing guidance for the prediction of the physical field. At the same time, the complex oscillation behavior of the keyhole is affected by the heat source input. The use of physical information input means provides a random fluctuation feature for the neural network prediction model, laying the foundation for accurately restoring the evolution law of the laser welding physical field. In summary, from the actual prediction results and theoretical analysis, the technical solution of the present invention has a solid theoretical basis, and the technical effect is consistent with that of the traditional method.

[0161] It should be noted that the embodiments of the present invention can be implemented by hardware, software, or a combination of software and hardware. The hardware portion can be implemented using dedicated logic; the software portion can be stored in a memory and executed by an appropriate instruction execution system, such as a microprocessor or dedicated design hardware. Those skilled in the art will appreciate that the above-mentioned devices and methods can be implemented using computer-executable instructions and / or contained in processor control code, for example, such as a carrier medium such as a disk, CD or DVD-ROM, a programmable memory such as a read-only memory (firmware), or a data carrier such as an optical or electronic signal carrier. The devices and modules of the present invention can be implemented by hardware circuits such as very large-scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, or programmable hardware devices such as field programmable gate arrays, programmable logic devices, etc., can also be implemented by software executed by various types of processors, or can be implemented by a combination of the above-mentioned hardware circuits and software, such as firmware.

[0162] The above description is only a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications, equivalent substitutions and improvements made by any technician familiar with this technical field within the technical scope disclosed by the present invention and within the spirit and principles of the present invention should be covered by the scope of protection of the present invention.

Claims

1. A high-precision and high-efficiency calculation method for laser welding numerical model based on physical information neural network, characterized in that: The steps include: Step 1: Generate training data based on the laser welding multiphase flow numerical model; Step 2: Build a laser welding physical field prediction model based on a three-dimensional convolutional long short-term memory network and introduce laser energy input information into the hidden layer; Step 3: Design a composite loss function that includes the residual of the physical control equation, data error, boundary condition error, and initial condition error; Step 4: Output the prediction results and use the gradient descent algorithm to minimize the loss function to adjust the network weights; Step 5: Iteratively update the model input based on historical physical field data to achieve time series prediction of the laser welding physical field.

2. The method according to claim 1, wherein The training data in step 1 is obtained by laser welding multiphase flow simulation, which includes heat transfer coupling between solid phase, liquid phase and gas phase, multiple reflection heat source model and solid-liquid phase change behavior. The extracted physical field data includes temperature, three-dimensional velocity components and fluid pressure.

3. The method according to claim 1, wherein The three-dimensional convolutional long short-term memory neural network constructed in step 2 uses a three-dimensional convolution kernel to extract the spatiotemporal characteristics of the physical field data, and introduces energy input information consisting of gas-liquid interface distribution, laser intensity, and absorption coefficient calculation into the hidden layer to restore the non-steady-state characteristics of the laser welding process.

4. The method according to claim 3, wherein The gas-liquid interface distribution is obtained by solving the conservation equation of the fluid volume fraction changing with time. The energy input intensity is calculated based on the light intensity at the reflection point, where the light intensity value is the product of the laser power and the Gaussian distribution function. The absorption coefficient is calculated based on the Fresnel formula and the normal direction of the reflection interface.

5. The method according to claim 1, wherein The physical information error term described in step 3 is composed of the residuals of the control equations of conservation of mass, conservation of momentum, and conservation of energy, and is used to ensure the physical consistency of the prediction results. The weights of each error term in the loss function can be adjusted to optimize the training process.

6. The method according to claim 1, wherein In step five, a sliding window method is used to reconstruct the input data. That is, the physical field data of the historical T moments are predicted for the subsequent T' moments. The input of the next T-T' moments is then concatenated with the predicted results as the input for the next round, so as to realize the recursive prediction of the physical evolution within the entire weld range.

7. A laser welding numerical modeling and prediction system for implementing the method according to any one of claims 1 to 6, characterized in that: include: A data preparation module is used to generate physical field data for training using a low-precision laser welding multiphase flow model; A model building module for building a three-dimensional convolutional long short-term memory network model with energy information input; A training module for optimizing neural network parameters based on physical information errors and data errors; The prediction module is used to recursively predict the physical field evolution results based on the trained model.

8. The system according to claim 7, wherein: The training samples generated by the data preparation module contain physical field data under different laser powers, welding speeds, and material thermophysical properties. The data format is a six-dimensional tensor, which represents the batch size, time step, three-dimensional space coordinates, and number of channels.

9. The system according to claim 7, wherein: The model building module includes multiple three-dimensional convolutional long short-term memory layers, which are used to extract composite features of space and time, and add convolutional layers to restore the feature mapping into five types of physical quantities, namely temperature, three-directional velocity components and fluid pressure.

10. The system according to claim 7, wherein: The prediction module can recursively output the physical field prediction results of subsequent multiple time steps based on the known laser welding physical field data of multiple time steps, and can adapt to the dynamic laser welding process modeling requirements under variable parameter conditions.

Images