🔬Why There Is No "AlphaFold for Materials" — AI for Materials Discovery with Heather Kulik

What It Covers

MIT chemical engineering professor Heather Kulik explains why materials science lacks an AlphaFold equivalent, covering active learning for multi-objective optimization, LLM limitations in molecular design, the gap between ML potentials and experimental ground truth, and how academic researchers can differentiate from well-resourced industry labs.

Key Questions Answered

• •LLM Chemistry Limitations: Test any LLM's chemistry capability with a concrete constraint task — Kulik asks every updated model to design a 22-atom ligand binding to a transition metal via two nitrogen atoms. No model has succeeded. LLMs perform at Wikipedia-level chemistry but fail at precise molecular design tasks that expert chemists solve in seconds.
• •Multi-Objective Active Learning: When optimizing materials across seven simultaneous objectives — CO2 selectivity, cost, aqueous stability, mechanical stability, thermal stability, and more — even low-accuracy ML models deliver 100x to 1,000x speed improvements per optimization dimension. The strategy is to begin optimization before models reach high accuracy, not wait for perfect models first.
• •ML Potentials Reliability Gap: Foundation models for interatomic potentials frequently fail outside their training distribution — molecules fall apart, predictions become unphysical. One high-profile 2024 model runs only five times faster than GPU-accelerated DFT calculations and produces unreliable results. Researchers should demand rigorous benchmarks against experimental data before replacing physics-based modeling with neural network potentials.
• •Literature Data Extraction Pitfall: When extracting material properties from published papers using LLMs, the numerical value reported in a graph and the author's written interpretation of that same graph frequently disagree. Teams building training datasets from literature must budget significant overhead for validation, as LLMs remain prone to false positives even with current models.
• •Academic Research Differentiation Strategy: With companies like Microsoft and Meta holding effectively unlimited compute, academic researchers should explicitly filter out problems solvable by brute-force scaling. Kulik's approach focuses on chemically complex, data-sparse domains — transition metal reactivity, excited-state behavior, processing-structure relationships — where domain expertise and creative problem framing outweigh raw computational resources.