Skip to content
← All projects
Research / ML for ScienceFirst-author manuscript in preparation

DKL-BO - Deep Kernel Bayesian Optimization

A CGCNN plus Gaussian-Process active-learning system for discovering high-band-gap 2D materials from a 3,351-crystal search space while evaluating only ~3% of candidates.

Public demo deferred until publication

Overview

DKL-BO is the strongest pure-ML portfolio asset: a Deep Kernel Learning surrogate paired with Bayesian optimization for pool-based materials discovery.

Because the manuscript is still in preparation, this page shows embargo-safe results, static plots, methodology, and implementation role without linking public code or a live demo yet.

Impact

  • Found the global best material within ~3% of evaluations, with 6.3x efficiency over random search.
  • Validated across 30 seeds and a second 10K-crystal SNUMAT study with paired Wilcoxon p approx 0.002.
  • Built a ~9.2K LOC Hydra pipeline with LMDB graph caching, pool-embedding reuse, calibration metrics, and 5 test suites.

What it shows

  • Graph neural networks, Gaussian Processes, uncertainty calibration, and acquisition-function design.
  • Research rigor through seed studies, statistical tests, transfer validation, and ablation-style sweeps.
  • Engineering depth through a configurable pipeline, graph caching, embedding reuse, and test coverage.

Demo plan

  • Before publication: static result page with convergence plots and summary tables.
  • After publication: Hugging Face Space showing a CSV-driven BO convergence animation.

Stack