Physics AI for Aerodynamic Databases
More than two months saved per design iteration, with trained Physics AI models generating aerodynamic databases in under an hour.
Business context
An aerospace engineering team needed aerodynamic databases to evaluate vehicle performance across operating conditions. Each design iteration depended on a new database generated with an established high-fidelity simulation workflow.
The challenge
Make aerodynamic database generation fast enough to support the customer’s engineering workflow, while validating model predictions against high-fidelity simulation results.
My role
I led the project end to end: CAD preparation, simulation execution, data-generation oversight, model training, inference code, customer integrations, and ongoing technical collaboration. I used Luminary’s existing machine learning infrastructure and worked with AI/ML engineers to improve model architectures.
Contributions & implementation
Geometry & Simulation: Built the CAD model and ran the high-fidelity CFD simulations used to generate training data.
Pipeline Delivery: Monitored data generation, debugged simulation-pipeline issues, and resolved delivery blockers throughout the project.
Model Training & Validation: Trained the Physics AI models and validated predictions against held-out simulation data, achieving less than 5% mean absolute percentage error on forces and approximately 10% error on moments.
Inference & Integration: Built the customer-facing inference code and integrations so the engineering team could generate aerodynamic databases using the trained model.
Iteration & Product Feedback: Worked directly with the customer to improve model performance, translated delivery feedback into product-engineering improvements, and collaborated with AI/ML engineers on more effective model architectures.
After an up-front simulation campaign and model training, the model generated an aerodynamic database in under one hour, saving more than two months per design iteration compared with the prior workflow.
Delivered the inference workflow for use in the customer’s production engineering organization, with continued iteration based on engineering feedback.
