An AI system proposes a material, predicts a useful property, and cites a paper. The explanation is convincing. Should a research team commit an experiment to it?
At Lymeric, our focus is advanced battery materials and material intellectual property. That makes the space between a promising idea and a justified experiment central to our work. Our collaboration with Ewha Womans University’s IMMS explores one part of that problem: how AI checks scientific claims.
We previously shared this research in an oral presentation at AI4Sci Korea 2026. Here, we take a closer look at the questions behind the work and what they mean for materials research.
A small window into a larger problem
The collaboration studies metal-organic frameworks, or MOFs: porous materials investigated for applications such as carbon capture and gas separation. Here, MOFs provide a setting for studying verification; they are distinct from our commercial focus on battery materials.
A claim about a material rarely stands alone. Its name has to identify the right structure. A reported result has to apply to the conditions being discussed. A relevant source may help without resolving the question.
Finding a relevant answer and establishing that it applies are different tasks.
Four questions before a verdict
MOF-Verify examines four kinds of verification: identifying the material, checking experimental conditions, assessing whether the evidence is sufficient, and recognizing when a question needs computation. Each asks something a fluent explanation can leave unresolved.
What supports the claim?
Are we discussing the right material?
A familiar name is only a starting point. The evidence must concern the material in the claim.
Does the evidence apply here?
A result under one set of conditions may not support a claim under another. The context matters alongside the result.
Is there enough to decide?
A relevant paper can still leave the question open. When the evidence cannot settle it, uncertainty should remain visible.
Does this need more than reading?
Some questions require a scientific calculation. A written explanation alone cannot settle them.
Better verification means a clearer next decision
The study reports improved verification performance across multiple language models compared with direct-answer and retrieval-based approaches. Those results concern defined research tasks. They do not demonstrate that a newly proposed material has been synthesized or is commercially ready.
The practical lesson is to preserve uncertainty where the evidence runs out. For a research team, “we need a different source” or “this requires a calculation” can be a useful outcome. It identifies the work still needed before a decision.
An experiment commits people, equipment, and laboratory time. The value of an AI proposal therefore includes how clearly it helps a team decide what deserves further investigation, what needs more evidence, and what can be set aside.
Three kinds of evidence. Three different questions.
A paper, a computational prediction, and a laboratory result each contribute something different. Keeping those roles visible helps a team understand what it knows—and which question remains open.
Different evidence, complementary roles
Literature
What has been reported?
Connect a claim to published observations and the context in which they were made.
Computation
What does a model predict?
Explore a question through a calculation, with the assumptions and limits of the model in view.
Experiment
What happens in the lab?
Observe how a material behaves under the conditions actually tested.
The next chapter: from verification to discovery
MOF-Verify focused on checking scientific claims. Building on that foundation, we are exploring how verification can help guide the search for materials with properties we want.
The starting point is what a material needs to do. The research direction is toward candidates worth investigating; the longer-term goal is to connect promising computational proposals with experimental evidence in the lab.
This work is underway. We’ll share more as the research develops.
The next question is taking shape: what should we make next?
Continue the conversation
Good science starts
with a good question.
Talk with us about materials research, discovery partnerships, or the ideas in this note.
Talk with Lymeric