On Challenges Using Long Short‐Term Memory Networks in Data‐Driven Inelasticity

ABSTRACT We present a framework that integrates long short‐term memory (LSTM) networks into a two‐dimensional data‐driven mechanics solver. We show that the staggered, double‐minimization algorithm induces solver‐generated noise, and that an LSTM trained on noise‐free stress–strain paths fails to account for these solver artifacts. To address this, we introduce two robustness strategies: (i) noise‐augmented training, where synthetic perturbations expand a library of 200 reference paths into 800 noisy variants and (ii) an attention‐enhanced approach, which embeds input‐ and output‐attention layers around the LSTM to dynamically extract and weight time‐step features. On a two‐dimensional plasticity benchmark, both approaches sharply reduce root‐mean‐square errors in stress and strain, with the noise‐augmented network achieving near‐perfect recovery of the reference paths and the attention‐augmented network exhibiting strong robustness without additional data. A remaining limitation is that the LSTM is trained on stress–strain histories generated by the same boundary value problem (BVP) on which we later deploy the data‐driven solver.