Research software
PhAST: evidence, scope, and next questions
PhAST is a public PyTorch-native research codebase for matrix-free phase-field fracture. This page is a short route through the published formulation, examples, and boundaries of the current evidence.
The problem
Differentiate a fracture workflow without hiding the mechanics
Phase-field fracture couples mechanics, damage evolution, irreversibility, and numerical solution choices. PhAST implements a matrix-free PyTorch route so that the relevant operations can support GPU execution and selected autograd-compatible analyses, while the finite-element mechanics remain explicit in the code and documentation.
FORMULATION
What is implemented
The public project documents phase-field fracture examples with explicit dynamics, damage solves, and inverse-analysis demonstrations. For equations, assumptions, and example configuration, use the documentation and associated preprint rather than this summary.
BENCHMARKS
What is demonstrated
The preprint presents dynamic and quasi-static benchmark material alongside gradient-based recovery of fracture energy. Each result remains tied to its stated mesh, hardware, precision, material model, and numerical settings.
INVERSE ANALYSIS
Why differentiability matters
Autograd-compatible routes make parameter recovery and related inverse questions practical to study. A useful inverse result still requires observability checks, gradient verification, and a forward model that is valid for the target experiment.
Interpretation
What the current record does not claim
- It does not claim a universal speedup over commercial or open-source solvers.
- It does not treat a learned proposal as a replacement for the mechanics or damage solver.
- It does not transfer a benchmark result to a new geometry, material, or loading path without validation.
Next evidence
What would make the case stronger
- Reproducible profiling on a declared hardware and numerical configuration.
- Matched validation against an independently configured reference problem.
- An inverse-analysis case with a documented observation model and noise sensitivity.