Abstract:Finite element simulations have become an integral part of product development and are extensively used in research as well. Depending on the complexity of the geometry or the physical problem, these simulations lead to a great degree of computational cost and time. In this study, we introduce a generative learning-based approach to accelerate finite element simulations. The aim is to explore the ability of generative adversarial networks to predict the output of a finite element simulation from a discretized domain with boundary conditions from a finite element pre-processor. The drawback of fully connected regression and recurrent models is that they are often bound to the distribution of training data and cannot extrapolate outside of the training domain. Generative models, however, can learn the underlying distribution of data and can efficiently extrapolate outside the training range. The goal here is to understand the ability of a GAN to replace FEM by incorporating the underlying mechanics knowledge into the network to enhance the generalizing ability of the network for various materials under various loading and boundary conditions. We propose an encoder–decoder-like U-Net generator architecture with activated skip connections and a PatchGAN-based discriminator that discriminates based on local patches of data, leading to a more robust critique for the generator. Both the generator and discriminator constitute convolutional blocks. The novelty of the proposed framework is that it takes inputs such as nodal information, their corresponding edges, nodal coordinates, and the boundary conditions for each particular node from a finite element pre-processor and computes the tensorial quantities of stresses and strains at each node as output that can be read by a finite element post-processor. Unlike the existing literature, where images are used as input and output. The proposed GAN architecture efficiently generates the solution to the given boundary value problem with an accuracy of about 97% and achieves an exact distribution of field variables as compared to a finite element solver. One drawback of using convolutional blocks inside the network is that they are not efficient at processing unstructured domains. To overcome this, we investigate the ability of graph neural networks to process the discretized domain and computing outputs and the future possibility of utilizing the graph architecture in the generator block.

Accelerated Solutions of Coupled Phase-Field Problems using Generative Adversarial Networks

LeapFrog: Getting the Jump on Multi-Scale Materials Simulations Using Machine Learning

Coupled Multiphysics Solver for Irregular Regions Based on Graph Neural Network

Towards General Neural Surrogate Solvers with Specialized Neural Accelerators

A Posteriori Evaluation of a Physics-Constrained Neural Ordinary Differential Equations Approach Coupled with CFD Solver for Modeling Stiff Chemical Kinetics

Learning two-phase microstructure evolution using neural operators and autoencoder architectures

A Physics Informed Neural Network (PINN) Methodology for Coupled Moving Boundary PDEs

A fully differentiable GNN-based PDE Solver: With Applications to Poisson and Navier-Stokes Equations

A generative learning and graph-based framework for computing field variables in Finite Element simulations

Mixed formulation of physics-informed neural networks for thermo-mechanically coupled systems and heterogeneous domains

Adaptive physics-informed neural networks for dynamic coupled thermo-mechanical problems in large-size-ratio functionally graded materials

Physics-informed MeshGraphNets (PI-MGNs): Neural finite element solvers for non-stationary and nonlinear simulations on arbitrary meshes

Solving Inverse Problems in Steady State Navier-Stokes Equations using Deep Neural Networks

Physics-Informed Graph-Mesh Networks for PDEs: A hybrid approach for complex problems

A mixed formulation for physics-informed neural networks as a potential solver for engineering problems in heterogeneous domains: Comparison with finite element method

Rethinking materials simulations: Blending direct numerical simulations with neural operators

Gradient Flow Based Phase-Field Modeling Using Separable Neural Networks

Dynamic and adaptive mesh-based graph neural network framework for simulating displacement and crack fields in phase field models

Graph and convolutional neural network coupling with a high-performance large-eddy simulation solver

A High Resolution Physics-informed Neural Networks for High-Dimensional Convection-Diffusion-reaction Equations.