NVIDIA launches Apollo open physics models

NVIDIA Apollo is a newly introduced family of open AI physics models built to bring real-time simulation speeds to industrial engineering and scientific computing.

Unveiled on November 17 at the SC25 conference in St. Louis, Apollo is designed to replace or accelerate conventional physics solvers that typically take hours or days to run, reducing turnaround times to just seconds.

What it does

Traditional engineering simulations rely on heavy numerical methods based on first principles. Apollo bypasses this bottleneck by providing AI surrogate models that learn physical laws and predict complex system behaviors nearly instantaneously while preserving engineering-grade accuracy.

The model suite tackles six core computational engineering domains: semiconductors and electronic design automation, structural mechanics, weather and climate forecasting, computational fluid dynamics (CFD), electromagnetics, and multiphysics such as plasma and nuclear fusion.

Major industrial players are already running Apollo models in production-grade environments. Applied Materials accelerated modules of its ACE+ multiphysics platform by up to 35x, using surrogate models to predict thermal, plasma, and flow dynamics in chip manufacturing chambers in seconds.

Cadence trained an AI model on high-fidelity simulation datasets to build a real-time digital twin of an entire aircraft. Siemens is baking Apollo into its flagship Simcenter STAR-CCM+ fluids software, and Synopsys reported speedups of up to 500x by initializing traditional fluid simulations with AI surrogates.

How it works and availability

Apollo achieves physical fidelity by combining state-of-the-art machine learning architectures, specifically neural operators, transformers, and diffusion methods, with domain-specific physics constraints.

Instead of treating systems as black boxes, the architecture is tailored to capture continuous operator mappings in partial differential equations, making it far more robust than standard generic neural networks.

NVIDIA is shipping Apollo with pretrained checkpoints alongside reference workflows for training, fine-tuning, inference, and benchmarking. Developers can either deploy the base checkpoints as plug-and-play surrogates or adapt them using their own proprietary simulation datasets.

The models will be hosted on Hugging Face, build.nvidia.com, and packaged as containerized NVIDIA NIM microservices.

Why it matters

If you work in CAE, aerodynamics, mechanical design, or semiconductor process modeling, Apollo is an obvious suite to test as soon as the weights land. Getting real-time inference during the design phase turns an agonizing multi-day optimization cycle into an interactive workflow.

Strategically, NVIDIA is cementing its hardware dominance in industrial engineering. Rather than competing with entrenched CAE vendors, it provides the open AI physics infrastructure that makes Siemens, Cadence, and Synopsys faster on NVIDIA GPUs. It is a smart, ecosystem-locking move that makes physical AI immediately practical for industrial labs.