Predicting Creep Failure by Machine Learning -- Which Features Matter?

TL;DR

This study evaluates the predictive power of spatial and temporal features for failure time prediction in disordered materials using machine learning on simulation data, highlighting the importance of strain over avalanche signatures.

Contribution

It demonstrates that global and local strain are more effective predictors of failure than avalanche statistics in simulated disordered material models.

Findings

Predictability increases with system size and temperature.

Avalanche features are not useful predictors in the ML models.

Strain measurements are key indicators for failure prediction.

Abstract

Spatial and temporal features are studied with respect to their predictive value for failure time prediction in subcritical failure with machine learning (ML). Data are generated from simulations of a novel, brittle random fuse model (RFM), as well as elasto-plastic finite element simulations (FEM) of a stochastic plasticity model with damage, both models considering stochastic thermally activated damage/failure processes in disordered materials. Fuse networks are generated with hierarchical and nonhierarchical architectures. Random forests - a specific ML algorithm - allow us to measure the feature importance through a feature's average error reduction. RFM simulation data are found to become more predictable with increasing system size and temperature. Increasing the load or the scatter in local materials properties has the opposite effect. Damage accumulation in these models proceeds in stochastic avalanches, and statistical signatures such as avalanche rate or magnitude have been discussed in the literature as predictors of incipient failure. However, in the present study such features proved of no measurable use to the ML models, which mostly rely on global or local strain for prediction. This suggests the strain as viable quantity to monitor in future experimental studies as it is accessible via digital image correlation.