DeepFEA: Deep Learning for Prediction of Transient Finite Element

Finite Element Analysis (FEA) enables the simulation of physical phenomena under various conditions. This is usually a computationally expensive and time-consuming process since it involves the numerical solution of partial differential equations. To accelerate FEA, effective surrogate methods based on Artificial Neural Networks (ANNs) have been recently proposed. However, these methods still have several limitations, mainly with respect to the dynamic prediction of accurate solutions independently from the original finite element model and/or respective ground truth data. This study proposes a deep learning-based framework for FEA (DeepFEA) that copes with these limitations. It is based on a novel ANN architecture composed of a multilayer Convolutional Long Short-Term Memory (ConvLSTM) network branching into two parallel convolutional neural networks with tensor outputs used to infer predictions related to the nodes and elements of FEA models. The architecture of the proposed network is optimized using a novel adaptive learning algorithm, called Node-Element Loss Optimization (NELO), which minimizes the error occurring at both of its branches. DeepFEA relies only on the initial and boundary conditions, as well as the external load of the modeled structure, which are provided as input to predict the transient solutions of an entire FEA simulation. The experimental evaluation of DeepFEA is performed on three datasets in the context of structural mechanics, generated to serve as publicly available reference datasets. The results indicate that it can accurately predict the outcome of multi-timestep FEA simulations, while offering a significant solution acceleration of at least two orders of magnitude.