Neural Network-Based Surrogate Modeling for Fast Solid Mechanics Simulations

Solutions of solid mechanics problems are available with reasonable accuracy through finite element analyses. However, performance studies with larger magnitude and real-time applications aren’t computationally viable. This paper proposes a compact neural network surrogate framework that easily learns the material and geometric parameters to nodal displacement mapping within the set-up a Multi-Layer Perceptron (MLP). The datasets from the ONSAS.m solver are parameter sweeps of two benchmark problems (a uniaxial prism and a composite cantilever). Our surrogate trained with domain RMSE is lower and it is significantly faster by orders of magnitude. We also verify extrapolation performance and visually compare displacements produced by the model with the ones from FEM. To sum up, we provide suggestions for consistency and uncertainty quantification for physics through operator learning and physicsinformed models.