AI Models Trained on Physics Are Changing Engineering

2026-04-09 · Source: IEEE Spectrum · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Emerging Technologies & Innovation · Depth: Intermediate, medium

Summary

Large physics models are emerging as a transformative technology in design engineering, mirroring the impact of large language models in software. These AI tools are beginning to amend or replace traditional physics simulations in industries like automotive and aerospace, significantly accelerating design workflows. For instance, General Motors has integrated an in-house large physics model to predict a car's coefficient of drag in minutes, a task that previously took two weeks of simulation. Companies like Neural Concept and PhysicsX are at the forefront, with PhysicsX collaborating with Nvidia to advance open standards for these models. While accuracy concerns exist for late-stage certification, early design phases benefit immensely from the speed, with some models demonstrating accuracy potentially surpassing simulations when fine-tuned with experimental data. Training approaches vary, utilizing architectures like transformers, geometric deep learning, or neural operators, and scaling laws observed in LLMs are now appearing in large physics models, suggesting improved generalizability and emergent properties as they grow.

Key takeaway

For AI Scientists developing engineering tools, consider integrating large physics models into design workflows to dramatically reduce simulation time. Your focus should be on training these models with existing simulation data and fine-tuning them with experimental measurements to achieve high accuracy for early design iterations. This approach empowers engineers to explore more design possibilities and make faster decisions, shifting their focus from low-value tasks to critical design choices.

Key insights

Large physics models drastically accelerate engineering design by replacing or augmenting traditional physics simulations.

Principles

• AI inference is orders of magnitude faster than simulation.
• Experimental data can enhance AI model accuracy beyond simulation.
• Scaling laws apply to large physics models, improving performance.

Method

Train an AI model on physics simulation results and experimental data to predict physical properties from 3D designs, enabling rapid iteration in early design stages.

In practice

• Integrate AI models for rapid drag coefficient prediction in car design.
• Use AI to explore a wider range of design options quickly.
• Fine-tune models with experimental measurements for higher accuracy.

Topics

• Large Physics Models
• Design Engineering
• Physics Simulation
• Aerodynamics
• Neural Concept

Best for: AI Scientist, AI Engineer, Machine Learning Engineer, Research Scientist

Related on AIssential

• How physics AI is transforming the future of space engineering — Physics AI dramatically accelerates aerospace design and certification by enabling near-instant, accurate simulations.
• Ten Years of Trying to Learn from a Mesh That’s Too Large — Large CFD meshes pose representational challenges for surrogate models, requiring specialized architectures to capture localized physical features.
• Introducing physics AI at Mistral: the foundation for engineering acceleration. - Mistral AI — Physics AI models accelerate industrial engineering by predicting physical behavior from geometry in seconds, enabling vast design exploration.
• When Deep Learning Meets Physics — Integrating physical laws into neural networks is a crucial, under-recognized advancement for AI's real-world utility.
• PhysicsX Raises $300M to Scale Physics AI - Let's Data Science — PhysicsX's $300M Series C highlights growing investor confidence in industrial AI for accelerating engineering simulation.
• Accelerating science with AI and simulations — AI is at a second inflection point, integrating language and multi-modal reasoning to accelerate scientific discovery.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by IEEE Spectrum.