Cross-Domain Algorithmic Discovery with an AI Co-Scientist

Asked to predict crystal structures, an AI co-scientist reached into computer vision and brought back a state-of-the-art algorithm. Searching across generative-modeling ideas from many fields, the autonomous agent selected MaskGIT, a masked generative model for images , as worth adapting to crystals, a connection neither a vision nor a materials specialist would naturally make.

The resulting model, the Masked Generative Crystal Transformer (MaskGXT) (the X stands for crystal, after Xtal, the conventional shorthand), adapts MaskGIT from images to crystals and sets a new state of the art on crystal structure prediction. The agent found the idea, instantiated it, and refined it largely on its own; a few sparse, high-level human hints carried it from competitive to best-in-class. This post is about that discovery process and the model it produced.

A new state of the art

A single MaskGXT model sets a new state of the art on match rate, the standard crystal structure prediction metric for whether a generated structure matches the target, across the standard benchmarks. It also leads on METRe, a stricter, polymorph-aware metric.¹

Benchmark	Metric	MaskGXT	Best baseline
MP-20	match rate ↑	73.79%	69.83%
MPTS-52	match rate ↑	36.75%	28.77%
MP-20	METRe ↑	74.78%	70.45%
MP-20 polymorph split	METRe ↑	79.06%	70.87%

Crystal structure prediction (CSP) asks: given a chemical composition, what stable crystal does it form? A crystal is a pattern of atoms repeated periodically in space, specified by three things: the atom types, a lattice (the unit cell that tiles space), and the fractional coordinates of each atom, which give its position as a fraction along the three lattice vectors so each value runs from 0 to 1 within the cell.

This representation raises several challenges. The coordinates need high precision to land on a stable configuration, and should follow the symmetric structure of a real crystal, captured by its space group. The model should ideally be invariant to periodic translation, since shifting the unit-cell origin leaves the crystal unchanged, and to atom ordering, since the atoms have no canonical order.

To handle these challenges, most recent CSP methods use continuous diffusion or flow models that denoise coordinates. MaskGXT takes a discrete route instead: it quantizes the crystal into tokens and predicts them, then refines each discrete prediction with a small continuous offset to recover the precision a coarse bin would lose. That route originates outside the field.

How it was discovered

The decisive move was a deliberate cross-domain transfer step. Rather than tune known CSP methods, the agent searched the broader generative-modeling literature for a framework not yet applied to crystals with a credible mechanism for the problem. It ran this as a tree-structured search: each node is one candidate model that the agent coded, trained under a fixed budget, scored on validation METRe, and used to decide what to try next. Over the course of the search, the agent proposed fourteen cross-domain generative frameworks from across ML, such as an autoregressive transformer (language), a Masked Generative Transformer (vision), and a Mamba/SSM interpolant (sequence modeling). The agent then focused its effort on the most promising of these, the Masked Generative Transformer, building it into the core formulation of MaskGXT. Explore the full search tree below, or open it full screen to see every node.

Although the agent discovered the core formulation on its own, the path to state of the art required sparse human intervention. On its own the agent already reached a roughly 70% match rate on MP-20, competitive with the prior state of the art; the interventions below carried it the rest of the way. In each, we proposed a mechanism or an objective, pointed to the relevant prior work, and left the implementation to the agent. Four were decisive. Three supplied a mechanism, a piece of knowledge from crystal generative modeling the agent was missing:

• Symmetry tokens. To improve structural accuracy, we proposed encoding symmetry, citing DiffCSP++ and WyFormer.
• Symmetry-preserving augmentation. To exploit that symmetry in training, we proposed orbit permutation over a symmetry-based atom ordering, citing MCFlow.
• Stratified sampling. To cover the multiple polymorphs a composition can form, we proposed sampling in a non-i.i.d. way.

The fourth supplied an objective: we asked the agent to recover the precision lost to discretization, and it arrived at a continuous sub-bin offset on its own.

What MaskGXT is

Once the search concluded, we organized the code the agent had written into the core methodology we describe here.

The transferred core is a discrete masked formulation. MaskGXT represents a crystal as a sequence of discrete tokens and predicts the masked ones with a transformer, the way MaskGIT fills in masked image patches. Three components adapt this core to crystals. Tokenization maps the lattice, fractional coordinates, space group, and Wyckoff positions to tokens, encoding crystal symmetry explicitly through the space group and Wyckoff tokens, with coordinates quantized on a circle so that bins near 0 and 1 stay adjacent. Ordinal, circular label smoothing trains the coordinate stream to treat neighboring bins as close, respecting that circular geometry. Sub-bin refinement then predicts a small continuous offset within each bin, recovering the precision that discretization would otherwise lose.

The discrete formulation also brings two advantages a continuous model lacks. Because the output is a finite vocabulary, greedy decoding can unmask the highest-confidence tokens first and return the single most probable structure. And because the space group is itself a token, space group stratified sampling can fix it to each likely value and branch generation, so a single composition yields distinct polymorphs, the source of the large polymorph-coverage gain.

Why it worked

This search succeeded because CSP offers cheap, well-aligned validation: a candidate trains in hours and is scored by a metric that tracks the goal. That tight loop let the agent run fourteen cross-domain bets and roughly five hundred refinements without a human judging each one. MaskGXT is one worked example, and it points toward a shift in how such research gets done. The AI co-scientist carries out the search; the human sets the goal and intervenes at a few decisive moments. The loop is harder to close in domains with long training cycles, physical experiments, or weak proxy objectives; developing more accurate proxy evaluations and determining how to incorporate human steering in such settings remain open problems.

References

• Chang et al., 2022. MaskGIT: Masked Generative Image Transformer. CVPR.
• Jiao et al., 2024. Space Group Constrained Crystal Generation (DiffCSP++). ICLR.
• Kazeev et al., 2025. Wyckoff Transformer: Generation of Symmetric Crystals (WyFormer). ICML.
• Martirossyan et al., 2025. All That Structure Matches Does Not Glitter (METRe). NeurIPS.
• Seong et al., 2026. Multimodal Crystal Flow: Any-to-Any Modality Generation for Unified Crystal Modeling (MCFlow). ICML.