Abstract: Simulating the behavior of macroscale brittle materials is extremely computationally expensive if every microcrack is explicitly simulated since a very fine grid is required to resolve these cracks. Continuum scale models that lump these microcracks into elastic moduli are much less computationally expensive but are prone to neglecting the physics of the subgrid-scale microcracks. We develop an approach that leverages data from an expensive simulator (HOSS) that explicitly resolves the behavior of microcracks to build a machine learning model of the impact of the microcracks on the elastic moduli. We then combine this machine learning model with a continuum scale model (FLAG) to produce a model (which we call ML-FLAG) that provides accuracy comparable to HOSS but is four orders of magnitude faster than HOSS. We validate our ML-FLAG model by comparing it to a flyer plate experiment.
Introduction
Predicting how brittle materials like ceramics or high-strength metals fail under dynamic loading is a significant computational challenge. High-fidelity models, such as the Hybrid Optimization Software Suite (HOSS), provide excellent accuracy by explicitly resolving microcrack networks. However, these models require massive computational resources, making them impractical for engineering optimization. Fast continuum-scale models like FLAG are efficient but often miss critical subgrid physics, leading to inaccurate failure predictions. This research introduces a machine learning (ML) workflow that bridges these scales, providing HOSS-level accuracy at FLAG-level speeds.
Methodology and Experimental Setup
The core of this approach is a Long Short-Term Memory (LSTM) recurrent neural network trained to act as a micro-scale physics emulator. The researchers generated a dataset of 100 high-fidelity HOSS simulations based on a flyer plate impact experiment.
Key to the model's efficiency was dimensionality reduction. The authors discovered that the macro-scale material response is primarily driven by the length of the longest crack and the maximum tensile stress, rather than the entire complex crack network. By predicting only these critical variables, the LSTM could effectively inform a modified statistical effective moduli model within the FLAG continuum code, allowing it to "account" for micro-scale damage without explicitly simulating it.
Results and Validation
The resulting ML-FLAG model achieved a massive four orders of magnitude ($10^4$) speedup over the high-fidelity HOSS simulations. Despite this speed, it maintained high accuracy in predicting material failure.
Validation against experimental VISAR (Velocity Interferometer System of Any Reflector) data showed that ML-FLAG successfully captured the "pull-back" signal in velocity profiles—a distinct signature of internal spallation that standard continuum models typically fail to resolve. Furthermore, the model demonstrated robust generalization; an emulator trained on impacts at 0.721 km/s accurately predicted experimental results at 1.246 km/s, suggesting it captured fundamental material physics.
Conclusion
By leveraging machine learning to bridge the gap between micro-scale fracture physics and macro-scale continuum mechanics, this work enables rapid, high-fidelity simulations of brittle material failure. Future work will focus on extending this 2D framework to 3D environments and exploring more diverse loading conditions, such as shear and tearing modes.
Summary updated