Project field note
Materials Informatics
Uncertainty-aware ML for predicting the Vickers hardness of metal-ceramic nanocomposites: physics-informed features, leakage-aware validation, benchmarking on ~11k real materials, and conformal prediction intervals that admit when they break.
A pipeline from composition and processing to predicted hardness with calibrated uncertainty, built to screen composite candidates before committing lab time. Small literature-scale materials data is easy to fit and easy to fool yourself with, so the project is organized around honest evaluation rather than a headline accuracy number.
Features combine Magpie composition descriptors (matminer) with physics-informed terms from known metallurgy: Hall-Petch grain-size strengthening, Orowan particle strengthening, and thermal-expansion mismatch. Models are gradient-boosted trees and random forests with Hydra-configurable backends, and SHAP attribution confirms the physics terms drive the predictions.
Evaluation is leakage-aware: paper-grouped cross-validation with a quantified gap against random splits, plus a volume-fraction extrapolation holdout. The same features and models are benchmarked on Matbench's elastic moduli tasks (~11,000 measured materials, official folds), landing mid-pack against the leaderboard as expected for composition-only features.
Every prediction carries a conformal interval implemented from scratch, with measured coverage: naive ensemble intervals under-cover badly, split conformal restores nominal coverage, and coverage honestly collapses under extrapolation. The hardness data itself is currently a small synthetic set; the methodology is validated on the real Matbench data, and real literature extraction is the next step.