AWS re:Invent 2025 - PhysicsX: Scaling Physics AI for Automotive Aerodynamics (STP109)

Summary of AWS re:Invent 2025 - PhysicsX: Scaling Physics AI for Automotive Aerodynamics (STP109)

Introduction to PhysicsX

• PhysicsX is an AI company focused on industrial applications, with a mission to "Build and deploy AI to accelerate industrial innovation"
• They have raised over $155 million in Series B funding and have a team of over 200 engineers, including AI researchers, simulation engineers, and software engineers
• PhysicsX partners with leading players across aerospace, defense, semiconductors, materials, energy, industrials, and automotive

The Evolution of Engineering Simulations

• Computational Fluid Dynamics (CFD) simulations, which approximate the Navier-Stokes equations, were a major step change from building physical prototypes
• While CFD simulations have fundamentally increased the speed of innovation, they are still a bottleneck, taking days to compute
• Physics AI models can perform the same simulations in a matter of seconds, transforming the speed at which manufacturers can innovate and design

Building AI Surrogates for Engineering Simulations

• AI surrogate models are trained on data from expensive physics simulations, allowing them to make fast, accurate predictions at a fraction of the computational cost
• The process involves:
  • Generating training data through expensive simulations
  • Storing and processing the data
  • Training, validating, and deploying the AI models
• PhysicsX has built a platform to support the entire AI lifecycle for engineering and physics, including simulation orchestration, model development, and a unified data backbone

Pre-Trained Models for Automotive Aerodynamics

• Existing public datasets for automotive aerodynamics, such as the DrivereT++ and Luminary datasets, have limited diversity, consisting of only a few baseline car models with morphed variations
• PhysicsX has trained attention-based architectures on these datasets, but found that the models performed poorly when tested on a more diverse "customer net" dataset
• To address this, PhysicsX is generating a larger, more diverse dataset of automotive aerodynamics simulations, with over 18,000 simulations and 120 baseline car models

Scaling Laws and Uncertainty Quantification

• As the size of the training dataset has grown, PhysicsX has observed improvements in the out-of-distribution performance of their models, measured by the mean absolute error of the pressure coefficient
• However, accuracy alone is not enough for engineering applications. PhysicsX's models also provide uncertainty quantification, allowing engineers to understand the confidence in the model's predictions
• This uncertainty quantification is crucial for active learning, where the model can identify the designs it is most uncertain about, and those designs can be used to further fine-tune the model

AWS Infrastructure for Scaling Physics AI

• PhysicsX leverages AWS infrastructure to power their physics simulations and model training:
  • Aerodynamic simulations are run on SageMaker Hyper Pod using GPU-based solvers
  • Simulation data is stored in S3, with pre-processing done using AWS Batch
  • Model training is performed on SageMaker Hyper Pod using data parallelism, with the file system hosted on FSx for Lustre
• The tight integration between the simulation and model training infrastructure is crucial, especially as PhysicsX looks to move towards training on transient simulation data

The Road Ahead

• PhysicsX will continue to innovate in automotive aerodynamics, including moving towards transient simulations
• They will also expand their work to other domains, such as radar cross-section, aeroelastics, and structural mechanics
• PhysicsX will explore both single, large-scope models as well as specialized models for different physics domains, working closely with customers to understand their needs

Key Takeaways

• Physics AI models can dramatically accelerate engineering simulations, transforming the speed of innovation
• Pre-trained models can help reduce the upfront data requirements for customers, but require careful attention to dataset diversity
• Scaling laws are emerging, where larger and more diverse datasets lead to better out-of-distribution performance
• Uncertainty quantification is crucial for engineering applications, enabling active learning and guiding model development
• Tight integration between simulation and model training infrastructure is key, especially as models move towards more complex transient simulations