Overview
At SAIC Volkswagen, I developed an active-learning framework for automotive aerodynamic performance prediction. It combined machine-learning models with CFD so that uncertain predictions could lead to new high-fidelity calculations and model updates.
The problem
A prediction model can provide a quick estimate, but its reliability depends on how well a new design resembles the available training data. The framework needed a way to identify cases requiring more simulation evidence.
Framework
The workflow combines MLP and GNN models with CFD and wind-tunnel information. It connects a design-stage prediction process to a training system that incorporates additional data.
Design stages, confidence checks, and the training loop in the aerodynamic prediction framework.
My contribution
I built the active-learning GNN–CFD workflow. Low-confidence predictions triggered high-fidelity CFD calculations, and the resulting data entered subsequent model retraining.
The objective was to direct simulation effort toward cases where the predictive model needed more information. This tied data-driven prediction to physical simulation rather than relying on the model alone.
Related work
I also worked on automated OpenFOAM workflows and aeroacoustic noise prediction during my time at Volkswagen.