🤖 machine learning

Discovering Crystal Structure Prediction Algorithms with an AI Co-Scientist

This paper introduces HACO, a human-AI co-discovery system that leverages cross-domain search and sparse human guidance to adapt the MaskGIT vision model into MaskGXT, a state-of-the-art algorithm for crystal structure prediction that achieves superior accuracy on standard benchmarks.

Original authors: Kiyoung Seong, Nayoung Kim, Sungsoo Ahn

Published 2026-06-23

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Kiyoung Seong, Nayoung Kim, Sungsoo Ahn

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to build a perfect LEGO castle, but you don't have the instruction manual. You only know the list of bricks you have (the chemical ingredients). Your goal is to snap them together in a way that creates a stable, beautiful structure. This is the challenge of Crystal Structure Prediction (CSP): figuring out how atoms arrange themselves into a solid crystal.

For a long time, scientists tried to solve this by writing their own computer programs to guess the arrangements. But in this paper, researchers from KAIST introduced a new way: they let an AI "Co-Scientist" figure out the best way to write the program itself.

Here is the story of how they did it, explained simply:

1. The AI Co-Scientist (HACO)

Think of the researchers as the "Project Managers" and the AI as a brilliant, tireless "Junior Engineer."

The Goal: The managers told the AI, "We need a program that can build stable crystals from chemical lists."
The Twist: Instead of telling the AI how to build it (e.g., "Use this specific math formula"), they gave it a huge library of ideas from other fields, like computer vision (how computers "see" images) and language processing.
The Search: The AI started testing thousands of different ideas. It was like a chef trying recipes from Italian, Japanese, and Mexican cuisines to see which one makes the best cake. Most ideas failed, but the AI kept trying, learning from every mistake.

2. The Big Discovery: Borrowing from Vision

The AI found a "secret sauce" in a completely different field: Computer Vision.
There is a technique called MaskGIT used to fix blurry or missing parts of images. Imagine a puzzle where some pieces are covered up; the AI looks at the surrounding pieces and guesses what's underneath.

The AI realized: "Hey, a crystal is just a 3D puzzle! If I treat the crystal like a picture with missing pieces, I can use this 'guessing' technique to build it."
This was a huge leap because no one had tried using this specific "image-fixing" method for building crystals before.

3. The Human Touch: Sparse Steering

The AI didn't work alone. The human scientists acted like Guides.

They didn't write code or tweak settings.
Instead, they gave high-level hints, like: "Hey, crystals have symmetry (they look the same if you rotate them), so make sure your guesses respect that," or "Sometimes the same ingredients can form different shapes (polymorphs), so try to find all of them."
The AI took these hints and rewrote its own code to include them.

4. The Result: MaskGXT

The final product is a new AI model called MaskGXT.

How it works: Instead of trying to draw the crystal continuously (like a painter blending colors), MaskGXT treats the crystal like a sentence of words. It breaks the crystal down into "tokens" (like Lego blocks representing the shape, the angles, and the symmetry).
The Process: It starts with a blank slate (all tokens masked/hidden) and fills in the blanks one by one, using confidence to pick the best "word" (or block) for each spot.
The "Polymorph" Trick: Since one set of ingredients can make different shapes, the AI doesn't just guess one shape. It looks at the "symmetry" clues first and branches out to try several different high-probability shapes at once, ensuring it doesn't miss any valid options.

5. Did it Work?

Yes, and it beat the competition.

The researchers tested MaskGXT on standard crystal databases (MP-20 and MPTS-52).
The Score: It matched the correct crystal structures 79% of the time when looking for different shapes (polymorphs), compared to about 71% for the previous best method.
It also got the best scores on standard tests, proving it is more accurate and better at finding the right shapes than older methods.

The Takeaway

This paper isn't just about a new crystal-building tool. It's a proof-of-concept that AI can help discover new scientific methods.

The AI looked outside its own field (borrowing from vision).
It combined that with human guidance (symmetry rules).
It built a better algorithm than humans could have easily designed on their own.

The researchers call this a "Human–AI Co-discovery." It shows that when humans set the goals and give high-level hints, and AI does the heavy lifting of searching and coding, they can solve hard scientific puzzles faster and better than before.

Technical Summary: Discovering Crystal Structure Prediction Algorithms with an AI Co-Scientist

Problem Statement
Crystal Structure Prediction (CSP) is a central challenge in computational materials discovery, requiring the generation of stable crystal structures from chemical compositions. The problem is constrained by periodicity (crystals repeat in space), fractional coordinates living on a torus, space group symmetry, and the existence of multiple polymorphs for a single composition. While recent advances in generative modeling (e.g., diffusion models, flow matching) have improved CSP, they often rely on continuous formulations that require stochastic or iterative sampling, lacking a finite discrete decision space for confident greedy decoding. Furthermore, existing AI research agents typically operate within fixed problem specifications, optimizing code or hyperparameters for a given model family rather than discovering new modeling principles across domains.

Methodology: The HACO System
The authors introduce HACO (Human–AI Co-discovery system), a framework designed to discover scientific algorithms through cross-domain search and sparse human steering. HACO operates on two core principles:

Cross-Domain Transfer: Instead of perturbing existing domain-specific models, the agent searches across generative modeling frameworks from diverse fields (vision, language, etc.) to identify transferable principles.
Sparse Human Interaction: Humans provide high-level objectives or mechanisms (e.g., "improve polymorph coverage" or "incorporate symmetry"), while the agent handles implementation, training, evaluation, and refinement.

HACO employs a tree-structured search orchestrated by an LLM (Claude Opus 4.7). The search proceeds in three stages:

Discovery Stage: Searches for the generative methodology (2-hour budget per candidate).
Scale-up Stage: Optimizes the selected methodology with larger budgets (12-hour budget).
Sampling Optimization: Tunes the sampling algorithm.

Key operators include idea (proposing cross-domain frameworks), draft (instantiating code), improve (making coherent changes), and debug. The idea operator specifically identified MaskGIT, a masked generative model from computer vision, as a promising candidate for CSP.

The Resulting Algorithm: MaskGXT
Through the HACO process, the system developed MaskGXT (Masked Generative Crystal Transformer). Unlike continuous diffusion models, MaskGXT formulates CSP as a discrete token generation task.

Tokenization: A crystal is represented as a sequence of $1 + 6 + 5N$ $1 + 6 + 5 N$ tokens:
- 1 Space Group token.
- 6 Lattice tokens (quantized lengths and angles).
- $5N$ site tokens: 3 quantized coordinate tokens, 1 Wyckoff position token, and 1 atom-type token.
- Atom types are provided as side information (conditioning).
Discretization & Smoothing: Continuous fractional coordinates are mapped to discrete bins on a circle (modulo 1). To preserve metric structure, the model uses ordinal, circular label smoothing for coordinates and standard ordinal smoothing for lattice parameters, rather than treating bins as unordered categories.
Symmetry Handling: The model learns to assign consistent space group and Wyckoff tokens across different unit cell origins and atom orderings via online augmentations (Euclidean-normalizer origin shifts and intra/inter-orbit permutations).
Architecture: A bidirectional Transformer trunk with QK-normalized self-attention and SwiGLU feedforward networks. It includes a lightweight regression head to predict sub-bin offsets, refining the geometric precision beyond the discrete bin resolution.
Sampling:
- Space Group Stratified Sampling: The model predicts a posterior over space groups. For polymorph coverage, it branches generation into separate streams for the highest-posterior space groups.
- Confidence-Ranked Greedy Decoding: Instead of stochastic sampling, the model deterministically unmasks tokens with the highest confidence (argmax), enabling efficient generation of high-probability structures.

Key Contributions

HACO Framework: Demonstrates that an AI co-scientist, guided by sparse human objectives, can identify a transferable modeling principle (MaskGIT from vision) and adapt it to a scientific domain (CSP) to produce a state-of-the-art algorithm.
MaskGXT: Introduces a discrete masked generative transformer for crystals that unifies lattice, coordinates, space groups, and Wyckoff positions into a single token sequence, enabling confident greedy decoding and explicit symmetry handling.
Performance Gains: Shows that a single MaskGXT model achieves state-of-the-art match rates on standard benchmarks and significantly improves polymorph coverage.

Experimental Results
The model was evaluated on the MP-20 and MPTS-52 benchmarks, as well as the MP-20 polymorph split (which tests coverage of multiple structures per composition).

Standard CSP (One-to-One): MaskGXT achieved the highest match rate (MR) on both MP-20 (73.79% unfiltered) and MPTS-52 (36.75% unfiltered), outperforming strong baselines like Crystalite, MCFlow, and FlowMM. It also achieved the lowest matched-pair RMSE.
Polymorph Coverage (METRe): On the MP-20 polymorph split, MaskGXT reached 79.06% Match-Everyone-to-Reference (METRe) accuracy, compared to 70.87% for the strongest baseline (Crystalite). This improvement is attributed to space group stratified sampling, which allocates generation budget to distinct high-probability space groups.
Ablation Studies:
- Removing symmetry tokens caused a significant drop in METRe on the polymorph split, confirming the necessity of explicit symmetry modeling for coverage.
- Removing sub-bin offset regression had minimal effect on match rate but substantially increased RMSE, proving its role in geometric precision.
- Greedy decoding improved single-structure accuracy, while stratified sampling was critical for polymorph coverage.

Significance and Claims
The paper claims that in domains characterized by "cheap, fast, and well-aligned validation" (like CSP with automated structure matching), transfer-guided interactive AI co-scientists can contribute to scientific algorithm discovery. Specifically, the system successfully:

Identified a modeling principle from an unrelated field (masked generative modeling in vision) and adapted it to the constraints of crystallography.
Combined this transfer with targeted human domain guidance (e.g., requesting symmetry-aware generation) to refine the algorithm.
Produced a competitive algorithm that advances standard benchmarks without requiring the human to specify low-level implementation details.

The authors note that the success of this loop relies on the specific structure of the CSP problem (fixed datasets, fast training cycles, and reliable validation metrics). They suggest that extending this approach to domains with physical experiments or weak proxy objectives would require more structured evaluation and stronger human steering.