Machine learning / uncertainty
ECC
Predicting materials, with uncertainty.
Research internship · UTRGV AIM LabConceptual illustration · not experimental data
My contribution
I grouped repeated tests to prevent composition leakage, trained target-specific quantile models, and connected them to support-constrained inverse design.
The question
Can we predict both strength and ductility of bendable concrete while representing uncertainty and avoiding leakage between repeated tests of the same composition?
What I built
I grouped 620 tensile tests into 276 unique mixes, then trained target-specific CatBoost quantile models for tensile stress and strain. Conformal adjustment calibrates the prediction intervals. A support-constrained NSGA-II inverse-design stage rejects requested high-strength and high-ductility combinations outside the observed mix domain.
Evaluation
At the nominal 80% interval level, observed coverage was 80.4% for stress and 82.6% for strain. Group-level R² was 0.749 and 0.605, respectively. These are reported evaluation results for this dataset; coverage on new material populations requires further validation.
For implementation and research details, explore the linked repositories and preprints.