Using data-driven machine learning (ML) models as surrogates in classical engineering is an emerging trend in the literature. However, effective surrogate modeling in path-dependent problems requires a deep understanding of the fundamental physical properties that naturally arise in data obtained from simulations or experiments. While generic ML architectures can capture nonlinear behavior, they may not inherently satisfy the specific temporal constraints dictated by physical processes. This study examines the characteristics of deformation paths generated through finite element simulations and identifies key modeling requirements for achieving physically meaningful predictions. One important requirement is that future inputs do not influence past outputs, a property typically satisfied by most surrogate ML models, yet rarely acknowledged or formalized. This requirement, often called the truncation condition, is essential for achieving physically meaningful predictions. Another closely related requirement is consistency across different time discretizations, which remains an active and important topic in deformation history modelling. To address these requirements, we propose a customized and adaptable Recurrent Neural Network (RNN) transition function that takes absolute strain inputs and is designed to enforce both truncation and consistency, ensuring robust predictions across varying temporal resolutions. This study contributes toward improving physically consistent damage initiation estimation and supports the development of more reliable surrogate models in computational mechanics.
Yatkın et al. (Sat,) studied this question.