Accelerating high-strain continuum-scale brittle fracture simulations with machine learning

M. Giselle Fernández-Godino, Nishant Panda, Daniel O’Malley, Kevin Larkin, Abigail Hunter, Raphael T. Haftka, G. Srinivasan

Abstract

Researchers have developed a machine learning workflow that accelerates high-strain brittle fracture simulations by four orders of magnitude ($10^4$). By employing a Long Short-Term Memory (LSTM) network to emulate micro-scale crack statistics, the team successfully bridged the gap between high-fidelity, computationally expensive models and fast, continuum-scale simulations. The resulting model matches experimental data even under novel impact conditions.

Introduction: The Challenge of Brittle Fracture

Simulating the failure of brittle materials like concrete, ceramics, or beryllium under high-speed impact is a significant computational challenge. High-fidelity models, such as the Hybrid Optimization Software Suite (HOSS), provide accuracy by tracking individual microcracks and their coalescence. However, these models are prohibitively slow, often requiring hundreds of processor hours for a single micro-scale simulation.

Conversely, continuum-scale models like FLAG are fast enough for engineering applications but typically "homogenize" damage. By treating the material as a uniform block, they often miss the critical subgrid physics of crack growth. This leads to inaccurate predictions of phenomena like "spallation," where internal damage causes a material to break apart due to reflected shock waves. The challenge lies in incorporating HOSS-level physics into a FLAG-scale timeframe.

Methodology: Bridging Scales with Machine Learning

The authors proposed a multi-step "upscaling" workflow designed to inject high-fidelity physics into continuum simulations.

Schematic of the flyer plate impact simulation setup, showing the flyer and target plates and the velocity tracer location.

The workflow consists of four primary stages:

  1. Data Generation: 100 high-fidelity HOSS simulations were conducted, modeling a beryllium flyer plate impacting a target.
  2. Dimensionality Reduction: Analysis revealed that the macro-scale response is dominated by the "longest crack" and maximum tensile stress. This finding allowed the researchers to simplify complex micro-scale data into a manageable set of features for ML training.
  3. LSTM Emulator: A Long Short-Term Memory (LSTM) network was trained on this reduced data. LSTMs are ideal for this task as they are designed to handle time-series data, predicting the evolution of crack statistics over hundreds of time steps.
  4. Integration: The ML-predicted statistics were integrated into the FLAG hydrodynamics code via a modified constitutive model, allowing the fast model to account for micro-scale damage without explicit simulation.

Results and Validation

The ML emulator achieved a massive speedup, running in seconds and making the entire workflow approximately 10,000 times faster than the original HOSS model.

Evolution of the damage field over time ($t=0$ to $1.2 \mu s$), illustrating the progression from initial micro-cracks to coalescence and spallation.

Beyond speed, the ML-informed FLAG model (FLAG-ML) demonstrated superior accuracy. It successfully captured the "pull-back" signal in shock wave velocity—a critical indicator of internal damage—which standard continuum models typically miss. Validation against experimental Velocity Interferometer System for Any Reflector (VISAR) data showed excellent agreement. Furthermore, the model exhibited robust transfer learning; when tested on impact velocities significantly higher than those in the training set (1.246 km/s vs. 0.721 km/s), it continued to perform accurately, suggesting it captured fundamental material physics rather than just memorizing the dataset.

Conclusion

While currently limited to 2D simulations and specific to Beryllium S200F, this workflow provides a code-agnostic template for accelerating complex fracture simulations. Future work will focus on extending the model to 3D environments—where crack branching is more complex—and testing its robustness across a wider variety of materials and loading conditions.

Summary updated

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.
BibTeX
@inproceedings{fernandezgodino2020accelerating,
  author = {M. Giselle Fernández-Godino and Nishant Panda and Daniel O’Malley and Kevin Larkin and Abigail Hunter and Raphael T. Haftka and G. Srinivasan},
  title = {{Accelerating high-strain continuum-scale brittle fracture simulations with machine learning}},
  booktitle = {Computational},
  year = {2020},
  doi = {10.1016/j.commatsci.2020.109959},
}

Loading related content...

Figure