A data-driven inelasticity framework enhanced by neural network based propagators – application to trusses

• We extend a data-driven inelasticity framework by Long Short-Term Memory neural network based propagators. • The proposed framework does not require specific and non-universal approaches for the history surrogate and should therefore be extendable to the prediction of other inelastic material behavior without any methodological changes. • A systematic and comprehensive hyperparameter study identifies the optimal architecture of the neural network for the material behavior under consideration. • The proposed framework achieves strain errors ≤ 0.02 % and stress errors ≤ 0.08 % for the simulation of complex truss structures. Research in computational mechanics is increasingly in search of methods that directly incorporate the information governed by data. Within these data-based methods, the data-driven mechanics approach seeks to replace conventional material models with discrete data sets. Offering immense potential in computational mechanics, where the simulation of complex mechanical systems is often challenging, the extension of the method to path-dependent material behavior is a crucial step. We achieved this extension by introducing history surrogates and an accompanying propagator which acts as an update rule. Finding suitable choices of these quantities can become challenging if they are selected by intuition. In the present paper, we explore the idea of utilizing a Long Short-Term Memory (LSTM) neural network as propagator, which automatically calculates and updates the history surrogate. In this context, the network is not deployed in order to represent the underlying material behavior by replacing the data, but rather to assist the data-driven solver, which continues to calculate the stress-strain prediction in accordance with the considered boundary value problem and thereby enforces equilibrium of forces and kinematic compatibility. We introduce the framework required to achieve accurate results and provide a thorough investigation of this framework, covering data generation, network architecture and network training. We demonstrate the capabilities of the LSTM enhancement for the example of plasticity with isotropic hardening and provide a comparison to a framework without the enhancement highlighting the improvements. Most notably, the proposed framework operates fully automated, without requiring user input, while maintaining the high accuracy of the approach without the improvements.