Markus Buehler on knowledge graphs for scientific discovery, isomorphic mappings, hypothesis generation, and graph reasoning

“If you read 1,000 papers and build a powerful representation, humans can interrogate, mine, ask questions, and even get the system to generate new hypotheses...

What you will learn

• Accelerating scientific discovery with generative knowledge extraction
• Understanding ontological knowledge graphs and their creation
• Transforming information into knowledge through AI systems
• The significance of ontological representations in various domains
• Visualizing knowledge graphs for human interpretation
• Utilizing isomorphic mapping to connect disparate concepts
• Enhancing human-AI collaboration for faster scientific breakthroughs”

The scientist interviewed here used generative LLMs to distil "a thousand papers... into some ontological knowledge graphs... changing information into knowledge... connect them to understand how ... [they] relate. ... ontological representations allow us to understand how these different building blocks... are related and how these relationships ultimately lead to certain properties... in science, we like to build models of such things..."

If "AI systems ... read a thousand papers and build a really, really powerful representation... humans can interrogate... or get the AI system to give us new hypotheses... And we can actually trace how the model thought about giving the answer" - which may make these systems less inscrutable.

The AI analyses each chunk of each paper, creating "1000s of different sorts of small graphs", which they then combine via their transitive properties, combining into one node the many nodes describing the same concept in different graphs.

Of course the massive dimensionality of the internal graph can't be visualised by humans. But they can explore a node's local neighbourhood, and ask an AI to identify how 2 concepts relate: "you want to build ... very scratch resistant paint ... using graphene... I want to understand how graphene is connected to crack-resistant coatings... find a path maybe between these concepts... the graph... will tell you how to get from graphene to scratch-resistant coatings"

If there's no connection, use isomorphic mapping: identify the graph structure in one graph, find a similar structure in another, and investigate: "I’m interested in how materials become resilient ... identify structures [in the graph] ... that describe ... resilience in materials" - now if can you find similar-looking structures in the other graph, they probably describe something like resilience.

Of course, you ask the AI to investigate this. They literally mapped "Beethoven’s Ninth... against some material structures... [finding] very similar graph structures that actually could be isomorphically mapped between music and the materials graph ... [where] nodes have things like adhesive force or beam and failure and characteristic length dimensions and structural features, buckling behavior... in the music world... things like tonality, the composer, then Beethoven, F major C major, different scales".

Then (I think) they built an agent to "try to explain how they could be ... connected... the answers actually were quite, quite, quite rich in the way they understood the topics... interpolating between these different domains ... because the graph gives the model something to think about" - without the graph any LLM's answer will be "not very rich, it’s not very nuanced... more generic".

These mappings are superuseful because they can be automated, and "because I might be an expert in one field, but not in the other", and because "if I understand, how these are related... I can maybe get a new hypothesis about it... now you’re beginning to discover new relationships in materials... inspired by music".

This could supercharge science: the interviewee builds on MIT's infinite corridor, which connects all MIT departments together and permits serendipitous discovery, as an example: "those random connections ... really is what makes discovery and innovation possible... What if AI could help us make these connections faster with more data? ... human-AI collaboration to come up with this interesting new idea ... But we need... the grounding in physics, in experimentation... and part of that can be automated as well ... multi-agent AI... [with] feedback from the world."

One more point: "isomorphisms... provide us with a structural substrate to say, Okay, here’s something there that’s universal ... in conventional science, we call this a foundational theory... not just true in one domain, but true in many different domains."

Also: they provided everything via an Apache 2.0 license.

Looking ahead, as LLMs improve: "In multi-agent AI, we were going to have multiple AIs talking to each other and communicating, if every one of them is just slightly better, the collective sum is going to be emergent."