physicsnemo: a scalable deep-learning framework for building andTraining physics-informed AI models for science and engineering

What it solves

PhysicsNeMo is a deep-learning framework designed for AI4Science and engineering. It solves the challenge of building, training, and deploying AI models that combine physical laws (physics-informed) with data-driven approaches to enable real-time predictions in domains like computational fluid dynamics (CFD), structural mechanics, and electromagnetics.

How it works

Built on top of PyTorch, the framework provides a modular stack of components:

• Models: A library of optimized architectures including Neural Operators (FNO, DeepONet), Graph Neural Networks (GNNs), Diffusion models, and Physics-Informed Neural Networks (PINNs).
• Datapipes: Scalable pipelines specifically tuned for scientific data structures like meshes and point clouds.
• Distributed Computing: A module based on torch.distributed for scaling training from single GPUs to multi-node clusters.
• Symbolic PDE Utilities: Tools for defining partial differential equations (PDEs) via SymPy to compute physics-informed losses with automatic spatial derivatives.

Who it’s for

It is intended for SciML (Scientific Machine Learning) researchers, climate scientists, and engineering domain experts who need to perform high-fidelity simulations and develop generalizable AI models for science.

Highlights

• GPU Optimization: Highly optimized for NVIDIA GPUs to maximize training speed and scalability.
• PyTorch Integration: Seamlessly integrates with existing PyTorch workflows and the broader ecosystem.
• Model Zoo: Includes a wide array of pre-implemented state-of-the-art SciML architectures.
• Extensibility: Supports ONNX for deployment and provides Pythonic APIs for adding new geometries and constraints.