Discovering Crystal Structure Prediction Algorithms with an AI Co-Scientist

TL;DR

HACO system uses MaskGIT for crystal structure prediction, achieving 79.06% METRe accuracy.

cs.LG 🔴 Advanced 2026-06-22 44 views
Kiyoung Seong Nayoung Kim Sungsoo Ahn
AI crystal structure prediction generative model cross-domain algorithm discovery

Key Findings

Methodology

The HACO system conducts scientific algorithm discovery through cross-domain search and sparse human steering. It identifies MaskGIT, a vision-based generative model, as a promising framework for crystal structure prediction, forming the MaskGXT model. MaskGXT discretizes crystal structures into tokens and incorporates crystallographic symmetry tokens, space group stratified sampling, and sub-bin coordinate refinement to enhance prediction accuracy.

Key Results

  • MaskGXT achieves 79.06% METRe accuracy on the MP-20 polymorph split, significantly outperforming the strongest baseline of 70.87%.
  • It also attains the best match rate on standard MP-20 and MPTS-52 benchmarks.
  • By introducing crystallographic symmetry tokens and space group stratified sampling, MaskGXT excels in multi-candidate generation.

Significance

This study demonstrates that in domains with fast and well-aligned validation, AI co-scientists can facilitate scientific algorithm discovery by identifying transferable modeling principles and combining them with human domain guidance. This approach not only automates implementation but also introduces innovation in cross-domain methodologies.

Technical Contribution

MaskGXT provides a novel generative modeling approach by discretizing crystal structures into token sequences and predicting them with symmetry information. Compared to existing continuous geometry generation methods, MaskGXT offers significant advantages in discretization and symmetry handling.

Novelty

This is the first application of the MaskGIT model from the vision domain to crystal structure prediction, innovatively combining discrete generative models with crystallographic symmetry, providing a new prediction framework.

Limitations

  • MaskGXT may perform poorly with extremely complex crystal structures due to its reliance on discretization and symmetry information.
  • The model's performance on larger datasets remains unverified.

Future Work

Future research could explore the application of MaskGXT in other materials science domains and optimize its performance on large-scale datasets. Additionally, further study on integrating human guidance to enhance model adaptability is needed.

AI Executive Summary

Recent advances in AI have expanded its application in scientific research, yet challenges remain in crystal structure prediction. Traditional methods often rely on continuous geometry generation, struggling with complex symmetry and polymorph coverage issues.

This paper proposes a novel approach, the HACO system, which integrates the MaskGIT model from the vision domain. By introducing discretization and symmetry information, it achieves efficient crystal structure prediction. The MaskGXT model excels in MP-20 and MPTS-52 benchmarks, particularly achieving 79.06% METRe accuracy on the polymorph split.

This approach not only showcases the potential of AI in scientific algorithm discovery but also offers new insights for cross-domain applications. Future research could further optimize the model and explore its application in other domains.

Deep Analysis

Background

Crystal structure prediction is a core problem in materials science, involving the generation of stable crystal structures from chemical compositions. Traditional methods often rely on continuous geometry generation, such as diffusion models and stochastic interpolant frameworks, but face limitations in handling polymorphs and symmetry.

Core Problem

The core problem of crystal structure prediction is accurately generating crystal structures with diversity and symmetry. Existing methods struggle with complex symmetry and polymorph coverage, limiting their effectiveness in practical applications.

Innovation

This paper's innovation lies in applying the MaskGIT model from the vision domain to crystal structure prediction, forming the MaskGXT model. By discretizing crystal structures into token sequences and predicting them with symmetry information, MaskGXT excels in polymorph coverage and symmetry handling.

Methodology

  • �� Use the HACO system for cross-domain search, identifying the potential of the MaskGIT model.
  • �� Discretize crystal structures into token sequences, incorporating crystallographic symmetry tokens.
  • �� Employ space group stratified sampling and sub-bin coordinate refinement to enhance prediction accuracy.
  • �� Validate model performance on MP-20 and MPTS-52 benchmarks.

Experiments

Experiments used the MP-20 and MPTS-52 datasets, comparing MaskGXT with existing baselines. Key metrics included METRe accuracy, with critical hyperparameters like token count and sampling strategy. Ablation studies analyzed the contribution of each component.

Results

MaskGXT achieves 79.06% METRe accuracy on the MP-20 polymorph split, significantly outperforming the strongest baseline of 70.87%. It also attains the best match rate on standard benchmarks, demonstrating its advantage in multi-candidate generation.

Applications

MaskGXT can be used for crystal structure prediction in materials science, particularly in scenarios requiring symmetry and polymorph coverage. Its efficient generation capability could drive the discovery and design of new materials.

Limitations & Outlook

While MaskGXT performs well in multiple benchmarks, it may struggle with extremely complex crystal structures. Additionally, its performance on larger datasets remains unverified, requiring further optimization in future research.

Plain Language Accessible to non-experts

Imagine you're cooking in a kitchen. You have different ingredients (chemical compositions) and need to combine them into a delicious dish (crystal structure). Traditional methods follow a recipe step by step, but sometimes the combination of ingredients is complex, like considering each ingredient's properties (symmetry and polymorphs).

The HACO system is like a smart chef assistant that can learn from other cuisines (cross-domain search) and adjust the cooking method based on your high-level instructions (human guidance). Ultimately, it helps you create a dish (crystal structure) that meets the requirements, more efficiently and tastily.

ELI14 Explained like you're 14

Imagine you're playing a game where you need to build a complex castle (crystal structure) using different blocks (chemical compositions). Traditional methods are like following a step-by-step instruction manual, but sometimes the manual isn't detailed enough, especially when building a particularly complex castle.

The HACO system is like a super-smart game assistant that can learn new building techniques from other games (cross-domain search) and adjust the building method based on your instructions (human guidance). In the end, it helps you build a more beautiful, complex castle, and faster too!

Glossary

MaskGIT

A generative model from the vision domain that generates data through masking and parallel decoding.

Used as the foundational framework for crystal structure prediction.

METRe

A metric for evaluating the accuracy of generative models, measuring the match between generated and reference structures.

Used to assess MaskGXT's performance in benchmarks.

HACO

A system combining high-level human guidance and AI cross-domain search for scientific algorithm discovery.

Used to identify suitable generative models for crystal structure prediction.

Crystallographic Symmetry

Symmetry features in crystal structures affecting their physical properties and generation process.

Modeled in MaskGXT through symmetry tokens.

Polymorph

Different crystal structures that the same chemical composition can form, with varying physical properties.

Covered in MaskGXT through space group stratified sampling.

Open Questions Unanswered questions from this research

  • 1 How can MaskGXT's performance be verified on larger datasets?
  • 2 How can the model be further optimized to handle extremely complex crystal structures?

Applications

Immediate Applications

Material Design

MaskGXT can be used to quickly predict new material crystal structures, aiding material scientists in designing more efficient materials.

Long-term Vision

Automated Scientific Discovery

By combining AI and human guidance, MaskGXT could drive automation in scientific discovery, transforming materials science research.

Abstract

We introduce Human-AI Co-discovery system (HACO) for scientific algorithm discovery through cross-domain search and sparse human steering. Starting from the goal of generating crystal structures from chemical compositions, HACO searched across generative modeling methodologies from multiple fields and identified MaskGIT, a masked generative model from vision, as a promising framework for crystal structure prediction (CSP). HACO instantiated this masked formulation as a discrete token model of crystal structure; guided by sparse high-level human objectives, it then added crystallographic symmetry tokens, space group stratified sampling for polymorph coverage, and sub-bin coordinate refinement, yielding the Masked Generative Crystal Transformer (MaskGXT). On the MP-20 polymorph split, MaskGXT reaches 79.06% match-everyone-to-reference (METRe) accuracy, compared with 70.87% for the strongest evaluated baseline. MaskGXT also attains the best match rate on standard MP-20 and MPTS-52 CSP benchmarks. These results provide evidence that, in domains offering cheap, fast, and well-aligned validation, transfer-guided interactive AI co-scientists can contribute to scientific algorithm discovery by identifying transferable modeling principles and combining them with targeted human domain guidance.

cs.LG cs.AI