DDD – Data-driven Materials Design
HT-MAX hypothesis: An NLP and AI-driven framework leveraging all available data sources will inform synthesis, characterization and modeling activities to accelerate materials discovery.
DDD Objectives
- Develop novel NLP-based methods to explore, survey, and ultimately catalogue and learn from the existing literature;
- Integrate physical constraints into ML models and quantify uncertainties from multifidelity data; and
- Apply new ML/UQ methods for active learning and Bayesian optimization to drive materials discovery.
UQ and Active Learning with Physics-Informed Multi-Fidelity ML: Michael Shields and Wei Chen
- Novel composite ML architectures will fuse multi-fidelity data sources through Bayesian neural networks (BNNs) for prediction of material properties and descriptors with uncertainty.
- “Explainable” Latent Variable Gaussian Process (LVGP) models will drive UQ-aware active learning-based materials composition and process design using Bayesian Optimization (BO).
- LVGP, BNNs, and other ML models will integrate physics through explicit constraints and manifold learning methods.
Natural Language Processing: Emma Strubell and Tony Rollett
- Build on previous success with annotation of abstracts and classification
- Develop a large language model for extracting data from papers.