Overview
Mechanical metamaterials derive their properties from geometric structure rather than material composition, making the design space enormous and difficult to explore manually. This project developed METASET, a method for automatically curating diverse, representative training datasets for machine learning models in metamaterials design.
Approach
The core insight is that shape diversity in the dataset directly controls the quality of the learned surrogate model. We used Laplace-Beltrami spectral descriptors as geometry fingerprints, then applied determinantal point process sampling to select maximally diverse subsets. This replaces ad-hoc data collection with a principled, automated pipeline.
Outcomes
Demonstrated that small, diversity-aware datasets outperform larger random ones for training property predictors. The method generalizes to any shape-parameterized design space. Published at ASME IDETC 2020, received Paper of Distinction.
Skills & Tools
Python, NumPy, scikit-learn, topology optimization, spectral geometry, Gaussian process regression