Key Moments
🔬There Is No AlphaFold for Materials — AI for Materials Discovery with Heather Kulik
Latent Space PodcastKey Moments
AI can now predict material properties up to 1000x faster, but these models often fail on complex chemistry and lack experimental validation.
Key Insights
AI was used to screen tens of thousands of materials, uncovering an unexpected chemical phenomenon that made polymer networks approximately four times tougher.
Metal-organic frameworks (MOFs), described as 'tinker toys' or 'Legos' for chemistry, offer precise chemical arrangements and have Nobel Prize-winning discoverers.
Current machine learning models are trained on 'boring' organic molecule data, lacking sufficient diversity for complex phenomena like transition metals or warm dense matter.
While AlphaFold revolutionized protein folding prediction with experimental benchmarks like CASP, materials science lacks a similar large-scale, experimental ground truth for AI validation.
LLMs can extract information from scientific literature, but their output often mismatches graph data due to author interpretation or errors, requiring human oversight.
Academic researchers must focus on creative problems solvable without massive computational resources, as large tech companies possess more compute power for brute-force approaches.
AI discovers tougher polymers through unexpected quantum phenomena
Researchers have successfully employed AI to accelerate the discovery of new materials, moving beyond merely speeding up existing computational models. In one significant demonstration, AI screened tens of thousands of potential materials, a process that would take months to years per experiment in a lab. This screening uncovered an unexpected chemical phenomenon leading to emergent properties in polymer networks, making the plastic about four times tougher. The AI-designed material was surprising to experimentalists, who would not have conceived of it independently, and it was successfully synthesized. This breakthrough has applications in increasing the lifespan and utility of plastics, addressing issues of durability and waste. The core of the discovery involved understanding how certain molecular components break apart to dissipate force, thereby toughening the material. While the general concept of 'controlled breaking points' for material strength was previously understood, the specific quantum mechanical stabilization mechanism identified by AI was novel for polymer materials.
The shift from cheminformatics to machine learning
Professor Heather Kulik's journey into data-driven materials discovery began early in her career, driven by impatience with the traditional, one-molecule-at-a-time approach. Initially, her group used AI to accelerate existing computational predictions. Around 2015-2016, Kulik recognized the growing importance of machine learning and rebranded their efforts, moving from 'cheminformatics' to 'machine learning.' A key early project involved a student who used neural networks for materials design, marking a significant leap into ML-based methods. This early work, initiated as a class project, laid the groundwork for more advanced applications of AI in molecular design.
Active learning for multi-objective material optimization
Active learning represents one of the most promising areas for AI in chemical sciences, particularly for tackling multi-dimensional design challenges. A current project involves optimizing metal-organic frameworks (MOFs) for direct air capture of CO2, requiring the simultaneous consideration of multiple objectives. These objectives include cost, stability in humid environments, selectivity for CO2 over other molecules, mechanical strength, and thermal stability – totaling seven different objectives in one active learning campaign. Even with less accurate ML models, active learning can achieve significant speedups, offering a thousandfold improvement for each dimension optimized. This approach is crucial for efficiently searching vast "haystack" spaces with multiple competing criteria, ensuring optimization begins even before models reach high accuracy.
The role and limitations of LLMs in chemistry
While Large Language Models (LLMs) like ChatGPT demonstrate impressive Wikipedia-level chemistry knowledge, they have significant limitations in practical molecular design. For instance, asking an LLM to design a ligand with a specific number of atoms (e.g., 22) and specific binding properties often results in failure to meet all constraints simultaneously. An expert chemist could typically perform such a task almost instantly. LLMs can be valuable for introducing concepts in unfamiliar areas or providing initial insights, and their performance has improved as more chemical knowledge is digitized. However, Kulik stresses the importance of chemists learning chemistry well enough to discern when LLMs are correct or incorrect. Relying blindly on LLMs without foundational knowledge is risky. While they can augment knowledge for areas not requiring a deep dive, they are best used as tools rather than complete replacements for expertise.
Data scarcity and complexity hinder AI in materials science
Significant challenges remain for AI in materials discovery due to insufficient and non-diverse datasets. Current ML models are often trained on "boring" chemistry, such as organic molecules and protein-binding data, neglecting more complex phenomena like transition metal chemistry, warm dense matter, or the behavior of materials under light excitation. This lack of data for complex physics and bonding types limits the robustness of AI predictions. For example, while repositories like Materials Project exist for crystalline materials, the data typically comes from lower-fidelity density functional theory (DFT) calculations, not experimental ground truth. This contrasts with protein folding, where community challenges like CASP provided experimental benchmarks that fueled breakthroughs like AlphaFold.
The quest for an 'AlphaFold for materials'
The development of a true 'AlphaFold for materials' faces hurdles due to the inherent complexity and variability of chemical bonding across the material space. While simple materials like aluminum are well-understood and can be modeled effectively by neural networks, more complex materials such as iron oxides or high-entropy alloys present greater challenges. A major problem is the lack of robust validation for these ML models, especially as they scale up to larger time and length scales. Unlike AlphaFold, which benefited from experimental data, many materials ML models lack this experimental ground truth. This makes it difficult to ascertain their correctness, and they can fail more catastrophically than even AlphaFold, which also has known failure modes. A more transparent approach is needed to assess whether these models can reliably replace conventional physics-based modeling, especially if they are to offer significant speedups (e.g., two orders of magnitude).
Integrating literature data and the challenge of interpretation
Integrating textual information from scientific literature into AI models, using Natural Language Processing (NLP) and now LLMs, can extract valuable property data. Kulik's group has used this to predict material properties like decomposition temperature. Interestingly, data extracted from graphs in papers often mismatches the authors' interpretations of those graphs, highlighting challenges in literature analysis. Obvious errors, author biases, and differing interpretations can lead to flawed training data for AI models. LLMs have improved literature extraction but remain sensitive to false positives, necessitating significant human effort to verify ingested data. This interpretation issue can bias discovery towards previously reported findings rather than novel ones.
Bridging the gap between computation, experiment, and industry
Bridging the gap between computational predictions and experimental reality is a critical bottleneck. While active learning and automated labs address synthesis speed, integrating processing into ML remains a nascent field. Academic researchers face a resource disparity compared to industry, especially in terms of computational power, necessitating a focus on creative, compute-efficient research questions. The traditional publication model, where data is embedded within papers, is inefficient for machine learning; there's a need for more systematic, machine-readable reporting. Ideally, a collaborative effort facilitated by initiatives like user facilities or public cloud labs, could provide computational researchers with access to automated experimental execution and ensure data is collected in a machine-learning-ready format from the outset. Furthermore, fostering shared facilities for high-throughput experimentation, potentially funded by government agencies or private foundations, could drive progress.
Mentioned in This Episode
â—ŹSoftware & Apps
â—ŹCompanies
â—ŹOrganizations
â—ŹBooks
â—ŹConcepts
â—ŹPeople Referenced
Common Questions
While ChatGPT is good for Wikipedia-level chemistry knowledge, it struggles with specific molecular design tasks, such as designing ligands with an exact number of atoms and specific binding properties, which an expert chemist can do quickly.
Topics
Mentioned in this video
A large language model discussed for its capabilities and limitations in understanding advanced chemistry concepts, particularly in molecular design.
A language model referenced in the context of learning chemistry and molecular design, highlighting its limitations for expert-level tasks.
A highly successful AI system for protein structure prediction, often used as a benchmark and point of comparison for AI in materials science.
Graphics Processing Unit, a hardware component used for accelerating computations, including DFT calculations.
A code developed by Heather Kulik's group for transition metal complex structure generation and metal-organic framework screening.
Massachusetts Institute of Technology, where Heather Kulik is a professor of chemical engineering.
A repository of DFT data for crystalline materials that provides leaderboards for evaluating computational models.
A project providing data and leaderboards for evaluating machine learning models in catalysis.
A US government agency supporting scientific research, which has initiatives related to cloud labs and high-throughput automation in materials science.
A pharmaceutical company where John Paul Jana is an assistant director running their inverse design program.
A technology company with significant computational resources that academic researchers contrast with their own limitations.
A technology company with substantial computational resources, similar to Microsoft, creating a disparity with academic research environments.
A platform where the 'mole simplifier' code is available for use and installation.
More from Latent Space
View all 199 summaries
65 minDreamer: the Agent OS for Everyone — David Singleton
88 minWhy Anthropic Thinks AI Should Have Its Own Computer — Felix Rieseberg of Claude Cowork/Code
35 min⚡️Monty: the ultrafast Python interpreter by Agents for Agents — Samuel Colvin, Pydantic
86 minNVIDIA's AI Engineers: Brev, Dynamo and Agent Inference at Planetary Scale and "Speed of Light"
Found this useful? Build your knowledge library
Get AI-powered summaries of any YouTube video, podcast, or article in seconds. Save them to your personal pods and access them anytime.
Try Summify free