Torch-DPhys: A Torch-Native, MOOSE-Compatible Framework for Differentiable Physics Simulation and Hybrid Machine Learning Integration

Key points

• Native tensor residuals support CPU and GPU execution, compilation of selected numerical kernels, and sparse direct linear solves.1
• Demonstrations cover plane-strain elasticity, one- and two-loop thermal-hydraulic transients, and an axisymmetric molten-salt fast reactor.1
• A residual-adjoint parameter-identification calculation recovers nominal pump head and heating power with relative errors below 0.06% from noiseless synthetic MOOSE observations.1
• These examples illustrate a foundation for gradient-based inference, control, and coupled physics-machine learning models in nuclear engineering.1

Sources (1)

• [1]Torch-DPhys: A Torch-Native, MOOSE-Compatible Framework for Differentiable Physics Simulation and Hybrid Machine Learning Integration
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 16, 10:52 PM

  We present Torch-DPhys, a Torch-native framework implemented in PyTorch for differentiable physics simulation and hybrid physics-machine learning.

  Native tensor residuals support CPU and GPU execution, compilation of selected numerical kernels, and sparse direct linear solves.

Extractive summary: sentences quoted from the sources.