Most important take away

There is no “AlphaFold for materials” because material science involves far more diverse chemical bonding and building blocks than the 20 amino acids in protein folding, and existing training data comes from low-fidelity computations rather than experimental ground truth. The real bottleneck is bridging computation to experiment: machine-learned potentials are not yet reliably faster or more accurate than physics-based methods across diverse chemistry, and experimental data at scale remains scarce.

Chapter Summaries

AI-Driven Polymer Discovery

Heather Kulik describes her group’s work screening tens of thousands of materials with AI to discover an unexpected quantum mechanical phenomenon in polymer networks. The AI-designed material was roughly four times tougher than conventional designs, surprising experimentalists who would never have arrived at the design on their own. The discovery has practical applications for plastic durability.

From Cheminformatics to Machine Learning

Kulik traces her path from studying individual molecules with quantum mechanical modeling to embracing data-driven methods. Around 2015-2016 she shifted from calling her work “cheminformatics” to “machine learning,” catalyzed by a student’s class project that became a foundational neural network paper for her group.

Active Learning for Multi-Objective Optimization

Active learning shines when optimizing across many objectives simultaneously. Kulik’s group is currently running a seven-objective active learning campaign for metal organic frameworks (MOFs) targeting direct air capture of CO2, considering cost, stability, selectivity, mechanical and thermal robustness. Each added dimension yields 100-1000x speedup over brute-force search.

Metal Organic Frameworks (MOFs)

MOFs, whose discoverers won the 2024 Nobel Prize in Chemistry, are like molecular LEGOs with infinite combinatorial possibilities. They are used for gas storage, sensing, separations, CO2 capture, catalysis, and drug delivery. Their precise chemical architecture enables targeted host-guest interactions.

Quantum Mechanical Modeling vs. Machine Learning

Traditional quantum mechanical predictions (DFT) can take hours to weeks per calculation. ML models can accelerate these, and Kulik’s group even uses neural networks on quantum wave functions to predict which computational method is best for a given material — a non-trivial mapping that defies simple heuristics.

Why There Is No AlphaFold for Materials

Unlike protein folding with 20 amino acids and the CASP benchmark grounded in experimental data, materials science has far more diverse building blocks, highly variable bonding, and training data from low-fidelity DFT rather than experiment. Machine-learned potentials that look impressive in benchmarks often fail on real problems — molecules fall apart, and some heralded models are only about 5x faster than GPU-accelerated DFT while being unreliable. There is no robust way to verify correctness at large scales.

LLMs and Chemistry Knowledge

LLMs are good at Wikipedia-level chemistry but fail at expert tasks like designing a ligand with exactly 22 atoms. Kulik recommends learning enough chemistry to assess when models are right or wrong, then using LLMs as augmentation tools rather than replacements.

Literature Extraction and Data Challenges

Kulik’s group uses NLP and LLMs to extract property data from published papers. A notable finding: the data extracted from graphs in papers often disagrees with what the authors claim about those same graphs. False positives from LLM extraction remain a significant overhead.

The Role of Academia vs. Industry

With companies like Microsoft and Meta having near-infinite compute, academics must focus on creative problems that cannot be solved by brute force alone. Kulik emphasizes pursuing problems that haven’t yet crossed industry’s radar.

Call to Action: MolSimplify

Kulik’s group develops MolSimplify (and MOFSimplify for MOFs), an open-source tool for transition metal complex structure generation and MOF screening, available on conda, GitHub, and as a web app. She welcomes feedback from users.

Summary

Heather Kulik, a professor of chemical engineering at MIT, discusses why AI for materials science is fundamentally harder than AI for protein structure prediction, and what actionable paths forward exist.

Key actionable insights:

• Active learning is the highest-leverage ML technique for materials discovery right now. Rather than waiting for models to be perfectly accurate, combining even rough ML predictions with multi-objective optimization yields 100-1000x speedups per dimension. Practitioners should prioritize active learning campaigns over perfecting single-point prediction accuracy.

• Do not trust machine-learned interatomic potentials blindly. Several high-profile foundation models for materials have been released with impressive benchmarks but fail on real chemistry — molecules fall apart, and speed gains over GPU-accelerated DFT are modest (around 5x, not the 100x often implied). Always validate against known stable structures before deploying these models on novel problems.

• The biggest gaps for ML engineers to fill are in reactivity prediction, diverse chemical bonding (especially transition metals), and excited-state phenomena. These areas lack benchmarks and leaderboards but represent some of the most impactful chemistry. Building datasets and benchmarks here would be high-impact contributions.

• Learn enough domain science to evaluate AI outputs. LLMs are unreliable for expert-level chemistry tasks (e.g., they cannot reliably design a molecule with a specified atom count). Career advice from Kulik: use AI to augment domain knowledge, not replace it. Starting from zero chemistry knowledge and relying on LLMs leads to undetected errors.

• Literature-extracted data is noisy. Authors’ interpretations of their own experimental graphs often disagree with the raw data. Anyone building ML models from literature-mined datasets should budget significant time for data validation and be aware of systematic biases toward previously reported results.

• Processing and manufacturing are underexplored. Predicting material properties is only part of the problem. How materials behave during manufacturing and at device scale is essentially unaddressed by ML — this is a wide-open research area.

• No specific stocks or investments were mentioned. However, the broader investment landscape was discussed: private startups in materials AI are attracting funding, and there is a noted gap in philanthropic and institutional funding for materials-for-climate compared to biotech.

• Tools to explore: MolSimplify and MOFSimplify (open source, available on GitHub, conda, and web) for transition metal complex and MOF design. Data repositories include Materials Project and Open Catalyst Project for benchmarking, and Kulik’s group has curated MOF stability datasets on their website.