physicsnemo

Why we included this project

PhysicsNeMo is aimed at engineers and researchers who want neural networks to respect physical laws, not just fit whatever data is handed to them. It is a Python framework built on PyTorch that covers physics-informed neural networks, neural operators, graph networks, and transformers, with composable modules for architectures, data pipelines that handle meshes and point clouds, and distributed training that can stretch from a single GPU to a cluster. The symbolic PDE layer stands out: you write the governing equations in SymPy and it produces physics-informed loss terms with automatic spatial derivatives, so you go from equations to a trainable model without hand-deriving every term. Teams working on CFD, weather and climate, and other AI4Science problems can also lean on companion modules for running inference on pretrained models. If you already use PyTorch, it is a practical way to get GPU-optimized, physics-aware training without assembling the plumbing yourself.