Pull to refresh
Logo
What Changed
Following Why Sign Up
MIT model pairs 3D crystal structures with text to help AI identify materials

MIT model pairs 3D crystal structures with text to help AI identify materials

New Capabilities

New system outperforms text-only AI at finding battery electrolyte candidates by 'seeing' atomic arrangements

October 18th, 2026: MIT News describes CLaC materials model

Overview

Updated Oct 23

MIT researchers have built a model called CLaC that reads a crystal's three-dimensional atomic arrangement alongside text descriptions. Their paper appeared in Nature Communications on August 26, and MIT News described the work on October 18.

Trained on 126,000 crystals from the Materials Project, CLaC's top-ranked crystal matched the text query 56.6% of the time in one retrieval test, but its crystal-system accuracy is about 34%. Asked for solid-state lithium-ion electrolytes, its top 30 candidates averaged a band gap of 3.23 electron volts, a wide gap that can block electron flow. The screen ranked by band gap, not ionic conductivity; Li6PClO5 and Li4SiO4 had each appeared in separate electrolyte work, so the list is a shortlist for testing, not proof of battery performance.

Why it matters

AI that reads atomic structure could screen battery and electronics materials in days, shortening discovery cycles that today run for years.

Questions about this story

Free account needed to ask — your question is kept and asked for you right after sign-up. Answers are public.

No questions yet — be the first to ask.

Key Indicators

95.4%
Zero-shot composition matching accuracy
CLaC trained on synthetic narratives matched structures to the right composition description at 95.4%, versus 17.8% for a Gemma text baseline.
34.0%
Crystal system classification accuracy
The model's weak point, roughly on par with its text baseline and tied to small batch size and imbalanced training data.
56.6%
Text-to-crystal retrieval accuracy
In the best configuration, CLaC's first-ranked crystal matched the text query 56.6% of the time, versus 37.4% for a comparison loss-function setup.
2
Top picks with independent electrolyte precedent
Li6PClO5 (ranked 4th) was named a promising candidate in separate work; Li4SiO4 (ranked 9th) had been synthesized as a stable solid-electrolyte thin film.
3.23 eV
Average band gap of top 30 candidates
The 30 highest-ranked lithium-containing candidates averaged 3.23 electron volts, a wide gap that can suppress electron flow in a solid electrolyte.

Voices

Curated perspectives — historical figures and your fellow readers.

Sign up to generate historical perspectives on this story.

Log in Sign Up

People Involved

JL
Ju Li
Carl Richard Soderberg Professor in Power Engineering, MIT Department of Materials Science and Engineering; senior author
Active researcher at MIT
 Elsa Olivetti
Elsa Olivetti
Jerry McAfee (1940) Professor in Engineering, MIT Department of Materials Science and Engineering; co-author
Active researcher at MIT
YP
Yang Jeong Park
First author; former MIT postdoc
Former MIT postdoc

Organizations Involved

MIT Department of Materials Science and Engineering
MIT Department of Materials Science and Engineering
Academic department
Host institution for CLaC research

The MIT department behind the CLaC model's development.
Massachusetts Institute of Technology (MIT)
Massachusetts Institute of Technology (MIT)
Research university
Institution publishing the research

The university whose labs produced CLaC.

Timeline

2011 October 2026

4 events Latest: October 18th, 2026 · 3 weeks ago
Tap a bar to jump to that date

  1. MIT News describes CLaC materials model

    Latest Research

    MIT News published an article on the CLaC model, whose paper appeared in Nature Communications in August 2026.

  2. CLaC paper published in Nature Communications

    Research

    The paper, "Scaling multimodal materials representation learning with synthetic narratives," appeared in Nature Communications.

  3. MatterChat structure-aware materials LLM published

    Research

    Berkeley Lab and partners introduce a multimodal LLM bridging atomic potentials with language models.

  4. Materials Project database launched

    Background

    Open computational database of materials properties seeded by the Materials Genome Initiative.

Scenarios

1

A CLaC-predicted material is synthesized and validated as a solid electrolyte

Likely Resolves by End of 2027

Discussed by: MIT researchers Ju Li and Elsa Olivetti; the Nature Communications paper's authors

Independent labs test CLaC's battery electrolyte candidates, synthesize one, and confirm it works. The approach then spreads to labs studying batteries, semiconductors, and catalysts, giving researchers a faster way to shortlist materials before expensive experiments.

2

CLaC's accuracy limits stall adoption in materials labs

Possible Resolves by Jan 31, 2028

Discussed by: The paper's own limitations section; ittechwire analysis

Crystal system classification stays weak at 34%, dataset biases persist, and real performance depends on defects, interfaces, and processing the ideal crystal model ignores. The community treats the model as promising but not yet ready to replace experimental screening.

3

Structure-aware AI consolidates into a unified materials discovery platform

Possible Resolves by Q2 2027

Discussed by: MatterChat developers (Berkeley Lab, NERSC); the trend toward multimodal materials AI

CLaC's structure-and-language alignment merges with MatterChat-style interatomic potential bridges and instruction following. The result is a platform that screens candidates, generates structures, and converses with researchers, becoming a standard interface for materials design.

Historical Context

2 moments from history that rhyme with this story — and how they unfolded.

November 2020

AlphaFold (2020)

DeepMind's AlphaFold won the CASP14 protein-folding competition, predicting protein 3D structures from amino acid sequences at near-experimental accuracy. The breakthrough gave biologists a way to see proteins' folded shapes without costly lab work.

Then

DeepMind released its predictions openly, and research groups worldwide began using them to study proteins within months.

Now

Structural AI became a foundational tool in biology; the successor AlphaFold2 later covered nearly all cataloged proteins.

Why this matters now

AlphaFold showed that giving AI 3D structural understanding of molecules transforms discovery. CLaC aims to do the same for inorganic crystal structures.

June 2011

Materials Genome Initiative (2011)

The Obama administration launched a federal program to accelerate materials discovery through computation, data sharing, and advanced tools. It funded the infrastructure that now hosts large crystal-structure databases.

Then

The initiative spurred growth of open computational databases like the Materials Project at Lawrence Berkeley National Laboratory.

Now

Those databases became the training bedrock for machine-learning materials models, including the data CLaC draws on.

Why this matters now

CLaC's training data comes from the Materials Project, a direct product of the Materials Genome Initiative. The model is the latest step in that program's data-sharing vision.

Sources

(9)
+3
MIT News ittechwire Nature Machine Intelligence NERSC AI Times CryptoFocus.nl 한국R&D신문 Insight Area Nature Communications