DeepFEA: Deep learning for prediction of transient finite element analysis solutions

Georgios Triantafyllou, Panagiotis G. Kalozoumis, George Dimas, Dimitris K. Iakovidis

Abstract: Finite Element Analysis (FEA) is a computationally intensive method for simulating physical phenomena. Recent advancements in machine learning have introduced surrogate models that can accelerate FEA. However, there are limitations in developing surrogates of transient FEA models that can handle dynamic input and multi-dimensional output. Motivated by this research gap, this study proposes DeepFEA, a deep learning-based framework for predicting the solutions of transient FEA simulations. Key contributions of this study include i) a multilayer Convolutional Long Short-Term Memory (ConvLSTM) network branching in two parallel convolutional neural networks, enabling the prediction of both node- and element-related output (displacement, stress and strain), ii) a novel adaptive learning algorithm, called Node-Element Loss Optimization (NELO) tailored for minimizing accumulated error produced by recursive predictions of the network, and iii) publicly available reference datasets generated by 2D and 3D FEA model simulations in the context of structural mechanics. Key study outcomes regarding DeepFEA include i) improved performance over relevant state-of-the-art methods; ii) less than 3 % normalized mean and root mean squared error; iii) inference times two orders of magnitude faster than traditional FEA; and iv) demonstration in a real-life biomedical application, which confirmed its effectiveness for accurate and efficient predictions of transient FEA simulations.