What is
AI-accelerated physics?
AI-accelerated physics means using a neural network to replace, or speed up, a slow physics calculation. Instead of solving the equations from scratch every time, the network has seen thousands of solved cases and predicts the answer directly, usually in a fraction of a second.
The idea
Traditional simulation (see the previous page) is accurate but slow: a single run can take hours. Many engineering questions, though, are variations on a theme: the same motor at a different load, the same car with a slightly different shape. A neural network can learn the pattern from a few hundred solved cases and then answer new variations almost instantly. The result is called a surrogate model.
The main approaches
Neural operators. Networks that learn the relationship between an input field (a shape, a boundary condition) and an output field (a pressure map, a temperature map). A 2024 review in Nature Reviews Physics reports that they can be "four to five orders of magnitude" faster than conventional simulators in fluid dynamics and weather forecasting.[6] The best-known one is the Fourier neural operator, introduced in 2020.[10]
Physics-informed networks. Instead of only learning from examples, the network is penalised during training whenever it breaks the physics equations. This needs less data and is useful when solved cases are scarce.[11]
Differentiable solvers. A different route: keep the physics equations, but write the solver so that a GPU runs it fast and so that it can be trained through, like a neural network. NVIDIA Warp does this, and it connects directly to the standard AI training tools.[5] This is the approach Invisible Force takes for its own physics modules.
Who makes the tools
| Tool | Maker | What it is |
|---|---|---|
| PhysicsNeMo (formerly Modulus) | NVIDIA, open source | A free toolkit with ready-made network designs for surrogates; version 2.2.2 as of September 2026.[1][2][3] |
| Ansys SimAI | Ansys (Synopsys) | Trains on 30 to 100 past simulations and predicts new ones in minutes.[7] |
| Simcenter PhysicsAI | Siemens (formerly Altair) | Deep learning on meshes; Siemens claims speed-ups of up to a thousand times.[8] |
| DoMINO | NVIDIA research | A neural operator that predicts airflow around cars from their 3D shape.[9] |
Speed-up figures on vendor pages are marketing claims. NVIDIA's developer page, for example, quotes 500 to 100,000 times depending on the case.[4] Treat them as upper bounds.
The catch
A surrogate knows only what it has seen. Ansys SimAI, for instance, needs dozens of conventional simulation runs before it can predict anything.[7] Outside its training range, it can be confidently wrong. In practice, AI-accelerated physics and traditional solvers are used together: the solver produces the truth for a set of cases, and the network makes that truth fast enough to run inside a robot simulator.
References
- NVIDIA. NVIDIA PhysicsNeMo (repository). September 2026.
- NVIDIA. nvidia-physicsnemo on PyPI. September 2026.
- NVIDIA. PhysicsNeMo documentation. 2026.
- NVIDIA. PhysicsNeMo developer page. 2026.
- NVIDIA. NVIDIA Warp documentation. August 2026.
- Azizzadenesheli et al., Nature Reviews Physics 6, 320-328. Neural operators for accelerating scientific simulations and design. April 2024.
- Ansys (Synopsys). Ansys SimAI product page. March 2026.
- Siemens. Simcenter PhysicsAI product page. 2026.
- Ranade et al., NVIDIA, arXiv 2501.13350. DoMINO: a decomposable multi-scale iterative neural operator for modeling large-scale engineering simulations. January 2025.
- Li et al., arXiv 2010.08895. Fourier Neural Operator for Parametric Partial Differential Equations (origin). October 2020.
- Raissi, Perdikaris and Karniadakis, arXiv 1711.10561. Physics Informed Deep Learning (origin of PINNs). November 2017.
Every link was opened and checked in September 2026. Links open in a new tab.