Peridynamics-enabled sequential learning for predicting nonlinear crack propagation

doi:10.1016/j.ijmecsci.2026.111970

Peridynamics-enabled sequential learning for predicting nonlinear crack propagation

2026-08-01

0

PRE

AI

DOI：10.1016/j.ijmecsci.2026.111970

Original

Original request for help

Save

Abstract

En 中文

Predicting nonlinear crack propagation remains a fundamental challenge in fracture mechanics due to strong path dependence, geometric nonlinearity, and sensitivity to evolving stress fields. Although peridynamics (PD) provides a physically consistent framework for fracture modeling, its high computational cost limits its applicability to large-scale and long-horizon problems. This study develops a PD-enabled sequential surrogate framework that combines PD-generated crack-tip trajectory data with long short-term memory (LSTM) networks for rapid crack trajectory prediction. The key innovation is to reformulate crack propagation as a trajectoryevolution problem rather than a conventional static mapping task. Crack-tip trajectories extracted from PD simulations are represented using equal-arc-length resampling and incremental displacement encoding, allowing the LSTM model to learn sequential correlations in PD-resolved crack path evolution. By leveraging sequential learning, the proposed approach captures temporal dependency and path-dependent behavior intrinsic to crack evolution. The framework accurately reproduces complex fracture phenomena, including post-branching trajectory continuation, large-angle turning, and mixed-mode propagation, with cumulative mean prediction errors below 8% of the total crack length across multiple benchmark problems. In addition, it achieves up to 3 & times; computational acceleration compared with full PD simulations. These results demonstrate that PD-enabled trajectory-continuation surrogate and sequential learning provides an effective and scalable paradigm for fracture prediction, highlighting the critical role of trajectory representation and temporal structure in bridging datadriven methods and computational mechanics.

Keywords:

Peridynamics

Machine learning

Crack trajectory prediction

Long short-term memory

Efficiency

Peridynamics-enabled sequential learning for predicting nonlinear crack propagation

Abstract

Journal

Organization

Cited Papers