A new arXiv preprint introduces Torch-DPhys, a PyTorch-native framework for differentiable physics simulation and hybrid physics–machine learning. The framework implements selected thermal-hydraulic modeling concepts and discretizations, though the abstract is truncated before listing the full set of supported models.
The key design choice is native integration with PyTorch, which allows gradients to flow through physics simulations and enables seamless coupling with neural networks. Its MOOSE compatibility suggests it can interoperate with existing simulation workflows, though the abstract does not detail the exact interface.
Because the source is a single abstract, the article is limited to the stated capabilities. No performance benchmarks, usage examples, or comparisons to other frameworks are provided in the available text.