ML Surrogate for Automotive Aeroacoustics & Noise-Aware Design Optimization

Project description

This six-month project develops a machine-learning surrogate model for automotive aeroacoustics using a provided dataset. The model will rapidly predict wind-noise quantities generated by exterior parts, initially focusing on the side-mirror and A-pillar region, including overall sound pressure level, noise spectra and side-glass pressure-fluctuation levels.

The surrogate is intended to provide near-instant evaluation compared with transient CFD and acoustic simulations. The project then extends to predicting acoustic impacts of part modifications and couples the acoustic surrogate with an existing drag-coefficient surrogate for noise-aware multi-objective design optimization.

Goals

• Preprocess the dataset, select a geometry representation and train a model predicting OASPL and noise spectra on held-out geometries using a leakage-free protocol.
• Evaluate MAE in dB, RMSE and correlation.
• Add uncertainty estimation and failure-mode analysis.
• Assess prediction of noise differences between baseline and modified parts, such as side-mirror variants.
• Attribute which regions drive predicted noise.
• Couple the acoustic and Cd surrogates to obtain a drag-versus-noise Pareto front and create a quiet-design gallery.

Required skills

Must-have: Python and ML fundamentals using PyTorch or TensorFlow; basic external-aerodynamics knowledge; acoustics basics including sound pressure level, dB, spectra/PSD and frequency bands; data preprocessing and experiment tracking; and ability to work with large datasets, I/O, storage and batching.

Nice-to-have: 3D learning, signal processing, aeroacoustics, multi-objective or Bayesian optimization, uncertainty and calibration, geometry-processing tools such as Open3D, trimesh, PyVista or meshio, and HPC/GPU training workflows.

Deliverables and schedule

Deliverables include an acoustic-surrogate package with preprocessing scripts, checkpoints, inference code and benchmarks; a part-level analysis module with uncertainty estimates, ΔdB predictions and attribution maps; a coupled Cd-plus-noise optimization loop with Pareto front and quiet-design gallery; and a clean reproducible repository, README, handover instructions and final report.

The suggested plan covers literature review and preprocessing in month 1, baseline OASPL modeling in month 2, model improvement and spectral prediction in month 3, part-swap and ΔdB analysis in month 4, optimization and attribution in month 5, and packaging, benchmarking, gallery and handover in month 6.

Monthly stipend with potential end-of-internship performance bonus and potential paper-publication co-authorship.