NVIDIA PhysicsNeMo (previously Modulus) is an open-source Python framework designed for building, training, and fine-tuning physics AI models at scale. It provides utilities that enable developers to build AI surrogate models combining physics-driven causality with simulation and observed data, enabling real-time predictions. From neural operators and graph neural networks (GNNs) to generative AI models, developers can create proprietary AI models to enhance engineering simulations and generate higher-fidelity data for scalable, responsive designs. PhysicsNeMo supports the creation and validation of large-scale digital twin models across various physics domains, from computational fluid dynamics and structural mechanics to electromagnetics.

Features

• PhysicsAI Model Architectures: Offers a variety of approaches tuned for training physics AI models, including physics-informed neural networks (PINNs), neural operators, GNNs, and generative AI-based diffusion models.
• Training State-of-the-Art Physics AI Models: Provides an end-to-end pipeline and utilities for training physics-ML models, from ingesting geometry to adding partial differential equations. Includes training recipes in the form of reference applications.
• Training at Scale: Provides GPU-accelerated distributed framework to build foundational scale models, supporting data parallel and model parallel training pipelines scaled to multi-node training.
• Physics AI Reference Pipelines: Offers diverse reference pipelines as starting points for developers to customize and build their own solutions, spanning use cases from computational fluid dynamics (CFD) and thermal analysis to climate and weather.

Benefits

• AI Toolkit for Physics: Quickly configure, build, and train AI models for physical systems in any domain, from engineering simulations to life sciences, with simple Python APIs.
• Customize Models: Download, build on, and customize state-of-the-art pretrained models from the NVIDIA NGC catalog.
• Near-Real-Time Inference: Deploy AI surrogate models as digital twins of your physical systems to simulate in near real time.
• Scale With NVIDIA AI: Leverage NVIDIA AI to scale training performance from a single GPU to multi-node implementations.
• Open-Source Design: Built on top of PyTorch and released under the Apache 2.0 license, providing the benefits of open source.
• Standardized: Work with best practices of AI development for physics-ML models, with an immediate focus on engineering applications.
• User Friendly: Boost productivity with user-comprehensible error messages and easy-to-program Pythonic API interfaces.
• High Quality: Use high-quality software with enterprise-grade development, tutorials for getting started, and robust validation and documentation.