Chemomechanical damage prediction from phase-field simulation video sequences using a deep-learning-based...

Area of Science:

• Materials Science
• Electrochemistry
• Computational Modeling

Background:

• Lithium-ion battery failure mechanisms are critical for wider adoption.
• Operando microscopy and phase-field models are used to study material heterogeneities and multi-physics coupling.
• Predictive models for imminent battery failure are underdeveloped.

Purpose of the Study:

• To explore convolutional long short-term memory networks for predicting damage in battery cathode materials.
• To use phase-field simulation videos as a proxy for operando microscopy data.
• To evaluate model performance using customized quantitative metrics.

Main Methods:

• Convolutional long short-term memory (LSTM) networks were applied to video sequences from phase-field simulations.
• Two models were trained: one using only damage videos, the other using damage and hydrostatic stress videos.
• Customized quantitative metrics were developed to assess model performance in predicting fracture behavior.

Main Results:

• Deep learning models demonstrated significant capability in predicting fracture behavior, including crack propagation angle and length.
• The models successfully predicted imminent failure using limited data from simulations.
• Combining damage and hydrostatic stress videos potentially enhanced predictive accuracy (further details in study).

Conclusions:

• Deep learning, specifically LSTM networks, offers a powerful tool for predicting battery material failure.
• This approach can forecast critical failure events like crack propagation in cathode materials.
• The findings pave the way for developing safer and more reliable lithium-ion batteries.