What 126,000 machine-written descriptions taught a crystal model

TL;DR

• MIT's Ju Li group published a Nature Communications paper aligning 3D crystal structures with language-model-written descriptions, and reports promising battery-electrolyte search results.
• The preprint trains a crystal-graph encoder and a BERT text encoder contrastively: true structure–text pairs score high, shuffled pairs score low.
• Human-written literature covers few crystals. Synthetic text covers many more, and that coverage is what makes the alignment learnable.

What happened

On October 18, MIT News reported a paper in Nature Communications titled "Scaling multimodal materials representation learning with synthetic narratives". It is led by Ju Li, with first author Yang Jeong Park (a former MIT postdoc) and co-authors Mayank Kumaran, Chia-Wei Hsu and Elsa Olivetti. The model pairs 3D atomic structures with natural-language descriptions, most of them generated by language models from computational databases. According to MIT News, the team trained two versions of the model, one on the AI-generated descriptions and one on text from papers only. The synthetic-data version "successfully learned how to align geometric structures with text descriptions while the other did not." Asked to find solid-state lithium-ion battery electrolytes, the model returned top results that "included materials that had been independently identified as promising candidates."

Those are the team's claims, relayed by MIT's own news office. I couldn't access the journal article itself, so the details below come from the February 2025 arXiv preprint by the same five authors. It is very likely an earlier version of the published work, but its title differs, and the journal version may use more data and report different numbers.

The paper in brief

CLIP-style models learn to match images and captions from hundreds of millions of examples. Crystals offer nothing close to that. The preprint notes that available crystal data runs to "just a few million" structures, and the literature concentrates on a small set of "hot" materials. CLaC (Contrastive Language–Crystals) fixes the text side by using 126k structure–text pairs, built from Materials Project crystals with descriptions previously generated by GPT-3.5. A graph neural network encodes each crystal, a SciBERT or MatSciBERT encoder reads the text, and a contrastive objective aligns the two. The best configuration retrieves the correct crystal for a text description 82.50% of the time from a pool of 1,024 candidates. It is notable as an early demonstration that zero-shot crystal–language matching works at all, and that synthetic text can make up for a thin, biased literature.

The math

Setup. The input is a dataset of pairs $(x_g, x_t)$: a crystal structure and a paragraph about it. The outputs are two encoders, $f_g$ with parameters $\theta_g$ and $f_t$ with parameters $\theta_t$. Their vectors $h_{\mathcal{G}}=f_g(x_g)$ and $h_t=f_t(x_t)$ should land close together when they describe the same material.

From a crystal to a graph

A crystal is a unit cell repeated forever. The cell is defined by three lattice vectors $\mathbf{a}_1,\mathbf{a}_2,\mathbf{a}_3$ and atoms at positions $\mathbf{r}_i$. The usual crystal-graph construction, introduced by CGCNN (one of the paper's two encoders), connects atom $i$ to every periodic image of atom $j$ that lies within a cutoff radius $r_c$:

$$E = \left\{ (i, j, \mathbf{n}) \;:\; \lVert \mathbf{r}_j + n_1\mathbf{a}_1 + n_2\mathbf{a}_2 + n_3\mathbf{a}_3 - \mathbf{r}_i \rVert < r_c,\ \ \mathbf{n}\in\mathbb{Z}^3 \right\}$$

Here $\mathbf{n}=(n_1,n_2,n_3)$ indexes which copy of the cell the neighbor sits in. The preprint uses Jarvis-tools' converter and only describes it as "neighborhood-based graph construction", so take this as the standard recipe rather than the paper's exact settings.

A thought experiment shows why the periodic images matter. Take a simple cubic crystal with one atom per cell and lattice constant $a$. With $a < r_c < \sqrt{2}\,a$, the graph has a single node and six edges, and every one of those edges connects the atom to a copy of itself one cell away. Ignore periodicity and the crystal disappears.

Node features start as CGCNN atom embeddings. Edge features are radial basis functions of the displacement vectors. A GNN (CGCNN or the equivariant PaiNN) then runs message passing: at each layer, every atom updates its vector from its neighbors' vectors and the connecting edge features. A pooling step averages everything into one vector $h_{\mathcal{G}}$, which is the paper's Eq. 2.

The CLIP recipe: InfoNCE

CLaC adapts CLIP, so start there. Take a batch of $N$ pairs, project both sides into a shared space, and compute cosine similarities $s_{ij}$ between crystal $i$ and text $j$:

$$\mathcal{L}_{\text{InfoNCE}} = -\frac{1}{2N}\sum_{i=1}^{N}\left[\log\frac{e^{s_{ii}/\tau}}{\sum_{j=1}^{N} e^{s_{ij}/\tau}} + \log\frac{e^{s_{ii}/\tau}}{\sum_{j=1}^{N} e^{s_{ji}/\tau}}\right]$$

Each term is an ordinary softmax cross-entropy whose "correct class" is the diagonal element: crystal $i$ should pick text $i$ out of the $N$ texts in the batch (first term), and text $i$ should pick crystal $i$ (second term). Every other item in the batch serves as a negative. The temperature $\tau$ is the same softmax temperature you know from classification.

Worked example. Take $N=3$. Crystal 1 has similarities $(0.8, 0.5, 0.3)$ to the three texts. At $\tau=1$, the softmax gives the true text probability $e^{0.8}/(e^{0.8}+e^{0.5}+e^{0.3}) \approx 0.43$, a loss of about 0.85. At $\tau=0.1$, the logits become $(8,5,3)$, the probability rises to about 0.95, and the loss falls to about 0.055. The similarity gap is the same in both cases. Temperature only controls how strongly that gap counts.

There is a known catch: $\log N - \mathcal{L}_{\text{InfoNCE}}$ is a lower bound on the mutual information between the two views, and it can never exceed $\log N$. CLaC trains with a batch size of 8, which caps the bound at $\log 8 \approx 2.08$ nats.

InfoNCE in a toy batch: one true match, N−1 equally close negatives

interactive

• InfoNCE loss
• MI estimate log N − loss
• ceiling log N

Hover or drag across the chart to read values.

similarity gap Δ (true minus negative)0.3temperature τ0.1

P(true match ranked first), N = 8:

74.2%

P(true match ranked first), N = 1,024:

1.93%

MI ceiling at batch size 8 (nats):

2.08

What CLaC actually optimizes

The preprint swaps InfoNCE for a Jensen–Shannon estimator taken from CLIP-Lite, which "does not depend on the number of negative samples". Rewriting its Eq. 5 with the logistic sigmoid $\sigma$, using $-\log(1+e^{-T})=\log\sigma(T)$ and $-\log(1+e^{T})=\log(1-\sigma(T))$, gives:

$$\hat{I}^{JSD}_{\omega}(h_{\mathcal{G}};h_t) = \mathbb{E}_{P(h_{\mathcal{G}},h_t)}\big[\log\sigma(T_\omega)\big] + \mathbb{E}_{P(h_{\mathcal{G}})P(h_t)}\big[\log\big(1-\sigma(T_\omega)\big)\big]$$

$T_\omega$ is a small discriminator network that scores a (graph, text) pair. The first expectation is over true pairs, drawn from the joint distribution. The second is over mismatched pairs, drawn from the product of the marginals, which in practice means shuffling within the batch. Maximizing this quantity is the same as minimizing binary cross-entropy for a logistic classifier that labels true pairs 1 and shuffled pairs 0. In other words, it is logistic regression on pairs. Each pair is scored on its own rather than competing in an $N$-way softmax, which is why batch size matters less. The name comes from the Jensen–Shannon divergence, a symmetric relative of the KL divergence in the paper's mutual-information definition (Eq. 4).

The full loss (Eq. 7) adds three more terms. Two are copies of the same objective applied within each modality, contrasting two augmented views of one crystal (edge removal, node dropping, subgraph sampling) and two masked views of one text. The fourth is a masked-language-model loss that keeps the BERT encoder from forgetting its pre-training. The graph encoder trains from scratch. The text encoder starts from pre-trained weights.

Zero-shot use. At test time, the model embeds a query, computes cosine similarity against every candidate from the other modality, and ranks them. Classification works the same way with a handful of candidate statements, such as different crystal systems, so it becomes multiple choice.

What the results say

• Retrieval: CLaC-PM (PaiNN + MatSciBERT) ranks the true crystal first 82.50% of the time in a pool of 1,024, and in the top 3 and top 10 95.91% and 99.37% of the time. Random top-1 would be about 0.1%. PaiNN beat CGCNN, and MatSciBERT beat SciBERT.
• Zero-shot classification: The model is strong on composition questions. The authors call structural understanding "not satisfactory". They blame the small batch, which yields more negatives that differ in elements than in crystal system, as well as GNN oversmoothing.
• Downstream language tasks: CLaC-PS reaches the highest average F1 (65.18) on the SOFC-slot named-entity benchmark, ahead of SciBERT and MatSciBERT. CLaC-PM gets 95.33% test accuracy classifying glass vs. non-glass abstracts, comparable to MatSciBERT.

Caveats. The retrieval test set is a held-out split of the same GPT-3.5 narratives, so it measures matching within the synthetic distribution, not against real literature. The authors excluded the papers-derived test split because its paragraphs often mention several materials. They acknowledge noise from GPT-3.5 and report Silhouette scores near 0 despite good scores on other clustering metrics. The model handles only ideal single crystals, with no defects, polycrystals or metal–organic frameworks. The abstract's claim of "surpassing latest large language models" rests on baselines described in supplementary material that is not part of the main text. Finally, the preprint trains with a discriminator but retrieves by cosine similarity, and its main text doesn't spell out how those two scores relate.

Connecting the dots

The preprint explains the main mechanism in the news story: two encoders, one shared embedding space, and a contrastive objective that rewards matched crystal–text pairs. It also explains the synthetic-vs-papers result MIT highlights. The papers-mined corpus, built from 4M papers, yielded 400k sentence pairs that covered only about 400 compounds. The GPT narratives cover roughly 300 times more crystals. A contrastive model learns only from negatives that differ in meaningful ways, so 400 heavily studied compounds give it little to contrast.

The news also goes beyond the preprint. The electrolyte-screening result has no counterpart in the arXiv version, which offers only a six-material illustration that matches LLZO with solid-state batteries. The title change to "Scaling…" also hints at a bigger dataset. Li's comparison to AlphaFold is an analogy, not a demonstrated result.

Why it matters

For practitioners, this offers a usable pattern for any scientific field where structures are plentiful and good descriptions are scarce: generate the descriptions from a database, then align them contrastively. For researchers, it raises a familiar question in a new domain. If a language model wrote the training text, how much of the model's "understanding" is the text generator's knowledge, errors included? The authors flag hallucination as a limitation themselves. Language-based screening could eventually point synthesis effort toward overlooked crystals, but only after it has been checked against experiment.

What to watch next: whether the published version tests retrieval on human-written text rather than synthetic text, and whether outside groups confirm the electrolyte candidates experimentally.

Go deeper

• The CLaC preprint on arXiv (PDF). The abstract page lists no code repository.
• MIT News: the announcement of the published paper
• The Nature Communications article, linked from MIT News. I couldn't open it, so check its numbers against the preprint's.
• Learning Transferable Visual Models From Natural Language Supervision (CLIP) by Radford et al., 2021: the image–text recipe CLaC adapts.