Computational Fluid Dynamics (CFD) Simulation

To get started developing a real-time digital twin, try out the NVIDIA Omniverse Blueprint for real-time computer-aided engineering digital twins.

This blueprint demonstrates a reference architecture for real-time digital twins by integrating AI-physics surrogates (trained using data from CUDA-X accelerated solvers) and interactive visualization. The blueprint demonstrates how to connect CFD solvers or AI surrogates to Omniverse through Universal Scene Description (OpenUSD), enabling real-time visualization of CFD simulation results. Developers can modularize components, such as swapping the PhysicsNeMo model for custom AI models, to tailor workflows to specific use cases.

Here are four features of the blueprint that can help developers start training and fine-tuning AI-physics models for faster CFD simulations with NVIDIA technologies. These features can be used as part of the overall blueprint or individually.

1. Train From Scratch or Fine-Tune Foundation Models

The blueprint demonstrates how to use a pretrained AI model from PhysicsNeMo, an open-source framework for training and deploying AI surrogate models using simulation data (for example, velocity and pressure fields). PhysicsNeMo supports hybrid training, combining CFD datasets with foundational models to reduce training time. In this particular blueprint, it uses the NVIDIA NIM™ microservice for the DoMINO model for automotive aerodynamics.

2. Build, Train, and Fine-Tune AI-Physics Models at Scale

Developers can use the NVIDIA PhysicsNeMo AI framework to embed governing partial differential equations (PDEs), for example, Navier-Stokes, into machine learning models like neural operators and graph neural networks (GNNs). The framework couples with CFD solvers to generate parametric training datasets and enforces physical laws via symbolic differentiation during training. The blueprint shows how PhysicsNeMo integrates with Omniverse to enable real-time feedback loops for digital twins, bridging simulation and operational decision-making.

3. Accelerate Simulations With NVIDIA Blackwell

The NVIDIA Blackwell architecture delivers the computational power needed for billion-cell simulations. The NVIDIA GB200 Grace™ Blackwell Superchip features NVIDIA NVLink™-C2C for ultra-high-bandwidth CPU-to-GPU communication. It enables CFD workflows to efficiently manage the complex data exchanges required for large-scale domain decomposition and ghost-cell updates. For example, with CUDA-X and NVIDIA Grace Blackwell, Cadence realized an over 48x speedup in a 10-billion-cell large-eddy simulation (LES) of a complete aircraft during takeoff and landing. The entire simulation ran on a single NVIDIA GB200-NVL72 system, doing the work of almost 300,000 CPU cores at 7x lower cost per simulation.

NVIDIA Blackwell GPUs also take advantage of CUDA®-aware Message Passing Interface (MPI) to optimize inter-GPU communication, delivering near-linear scaling even as simulation sizes grow dramatically. This directly translates into real-world impact. With NVIDIA Blackwell, engineers can perform high-fidelity, end-to-end CFD simulations, unlocking new possibilities for rapid design iteration, real-time digital twins, and operational analytics, without compromising accuracy or reliability.

4. Integrate the End-to-End Workflow

Developers can combine these technologies into unified pipelines, such as CAD → meshing → GPU-accelerated CFD solve → AI surrogate → Omniverse visualization. Leading ISVs like Ansys, Cadence, and Siemens are bringing these capabilities to their customers today. This integration accelerates time to insight while maintaining gold-standard accuracy, enabling rapid design exploration and real-time operational analytics—all within a CFD simulation software.