Prediction of creep failure time using machine learning.

Area of Science:

• Materials Science
• Physics
• Computational Science

Background:

• Subcritical loads on disordered materials induce creep damage, progressing through decelerating, steady-state, and accelerating creep regimes.
• Predicting material failure from creep rate has been challenging due to complex spatial correlations in damage.
• Acoustic emission patterns reflect damage accumulation but are difficult to interpret for failure prediction.

Purpose of the Study:

• To develop a more effective method for predicting the remaining time to failure in disordered materials.
• To leverage supervised machine learning to analyze complex spatio-temporal damage patterns.
• To improve the accuracy of failure prediction compared to existing approaches.

Main Methods:

• Utilized a mesoscale elastoplastic model to simulate creep damage evolution in disordered solids.
• Employed supervised machine learning algorithms to process temporal signals from the simulations.
• Analyzed time series data of acoustic emissions as a proxy for damage accumulation.

Main Results:

• Machine learning models successfully estimated the remaining time to failure.
• Predictability was higher in materials with greater disorder and lower in larger systems.
• The machine learning approach significantly outperformed previously proposed prediction methods.

Conclusions:

• Supervised machine learning offers a robust framework for predicting material failure from creep damage signals.
• The method's effectiveness is influenced by material disorder and system size.
• This advanced technique provides a substantial improvement in predicting catastrophic breakdown in disordered materials.