Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

LLM Meets Diffusion: A Hybrid Framework for Crystal Material Generation

Subhojyoti Khastagir, Kishalay Das, Pawan Goyal, Seung-Cheol Lee, Satadeep Bhattacharjee, Niloy Ganguly

TL;DR

This paper introduces CrysLLMGen, a hybrid framework that combines an autoregressive LLM with a diffusion model to generate crystal materials by jointly handling discrete atom types and continuous coordinates and lattice parameters. The LLM proposes a chemically informed atom composition while the diffusion model refines fractional coordinates and lattice to ensure structural validity and thermodynamic stability. Through extensive experiments on De Novo, S.U.N., and text-conditioned generation tasks, the method outperforms both pure LLM and pure diffusion baselines in structural and compositional validity and achieves higher stability and novelty, with strong conditional generation capabilities. The approach demonstrates a practical pathway to scalable, controllable crystal material design, and is architecture-agnostic to accommodate future advances in LLMs and diffusion models.

Abstract

Recent advances in generative modeling have shown significant promise in designing novel periodic crystal structures. Existing approaches typically rely on either large language models (LLMs) or equivariant denoising models, each with complementary strengths: LLMs excel at handling discrete atomic types but often struggle with continuous features such as atomic positions and lattice parameters, while denoising models are effective at modeling continuous variables but encounter difficulties in generating accurate atomic compositions. To bridge this gap, we propose CrysLLMGen, a hybrid framework that integrates an LLM with a diffusion model to leverage their complementary strengths for crystal material generation. During sampling, CrysLLMGen first employs a fine-tuned LLM to produce an intermediate representation of atom types, atomic coordinates, and lattice structure. While retaining the predicted atom types, it passes the atomic coordinates and lattice structure to a pre-trained equivariant diffusion model for refinement. Our framework outperforms state-of-the-art generative models across several benchmark tasks and datasets. Specifically, CrysLLMGen not only achieves a balanced performance in terms of structural and compositional validity but also generates more stable and novel materials compared to LLM-based and denoisingbased models Furthermore, CrysLLMGen exhibits strong conditional generation capabilities, effectively producing materials that satisfy user-defined constraints. Code is available at https://github.com/kdmsit/crysllmgen

LLM Meets Diffusion: A Hybrid Framework for Crystal Material Generation

TL;DR

This paper introduces CrysLLMGen, a hybrid framework that combines an autoregressive LLM with a diffusion model to generate crystal materials by jointly handling discrete atom types and continuous coordinates and lattice parameters. The LLM proposes a chemically informed atom composition while the diffusion model refines fractional coordinates and lattice to ensure structural validity and thermodynamic stability. Through extensive experiments on De Novo, S.U.N., and text-conditioned generation tasks, the method outperforms both pure LLM and pure diffusion baselines in structural and compositional validity and achieves higher stability and novelty, with strong conditional generation capabilities. The approach demonstrates a practical pathway to scalable, controllable crystal material design, and is architecture-agnostic to accommodate future advances in LLMs and diffusion models.

Abstract

Recent advances in generative modeling have shown significant promise in designing novel periodic crystal structures. Existing approaches typically rely on either large language models (LLMs) or equivariant denoising models, each with complementary strengths: LLMs excel at handling discrete atomic types but often struggle with continuous features such as atomic positions and lattice parameters, while denoising models are effective at modeling continuous variables but encounter difficulties in generating accurate atomic compositions. To bridge this gap, we propose CrysLLMGen, a hybrid framework that integrates an LLM with a diffusion model to leverage their complementary strengths for crystal material generation. During sampling, CrysLLMGen first employs a fine-tuned LLM to produce an intermediate representation of atom types, atomic coordinates, and lattice structure. While retaining the predicted atom types, it passes the atomic coordinates and lattice structure to a pre-trained equivariant diffusion model for refinement. Our framework outperforms state-of-the-art generative models across several benchmark tasks and datasets. Specifically, CrysLLMGen not only achieves a balanced performance in terms of structural and compositional validity but also generates more stable and novel materials compared to LLM-based and denoisingbased models Furthermore, CrysLLMGen exhibits strong conditional generation capabilities, effectively producing materials that satisfy user-defined constraints. Code is available at https://github.com/kdmsit/crysllmgen
Paper Structure (31 sections, 6 equations, 2 figures, 4 tables, 1 algorithm)

This paper contains 31 sections, 6 equations, 2 figures, 4 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work: Crystal Material Generation
  3. Preliminaries
  4. Crystal Structure Representation
  5. Symmetry in Crystal Structure
  6. Methodology
  7. Problem Formulation
  8. Proposed Framework: CrysLLMGen
  9. Large Language Model (LLM) $f^{LLM}_{\phi}$ for Crystal Materials
  10. Diffusion Model $f^{Diff}_\theta$ for Crystal Materials [Atom Coordinates, Lattice Structures]
  11. Training and Sampling
  12. Experiments
  13. De Novo Material Generation (Gen)
  14. Stable, Unique and Novel (S.U.N.) Materials
  15. Text Conditioned Generation
  16. ...and 16 more sections

Figures (2)

  • Figure 1: Model Architecture of our proposed CrysLLMGen. The sampling process begins with prompting the LLM to generate intermediate representations for $\textbf{A}$, $\textbf{X}$, and $\textbf{L}$. The predicted atom types ($\textbf{A}$) are retained as-is, while the atomic coordinates ($\textbf{X}$) and lattice parameters ($\textbf{L}$) are further refined through a denoising diffusion network.
  • Figure 2: (a) Histogram of $(E^{hull})$ distribution computed after relaxation with CHGNet.(b) Results of Text-Conditional Generation on atomic compositions and space group.