LLM Meets Diffusion: A Hybrid Framework for Crystal Material Generation
Subhojyoti Khastagir, Kishalay Das, Pawan Goyal, Seung-Cheol Lee, Satadeep Bhattacharjee, Niloy Ganguly · 2025
Read LLM Meets Diffusion: A Hybrid Framework for Crystal Material Generation by Subhojyoti Khastagir, Kishalay Das, Pawan Goyal, Seung-Cheol Lee, Satadeep Bhattacharjee, Niloy Ganguly (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2510-23040.txt.
Introduction
Subhojyoti Khastagir and five co-authors combine language models with diffusion models to design new crystal materials. Language models handle discrete atom types well but struggle with exact positions and lattice sizes. Denoising models do the reverse: they fit continuous values but often miss the right mix of atoms.
Their hybrid, CrysLLMGen, joins both strengths in one generator for periodic crystals. The language side proposes compositions while the diffusion side refines geometry. Materials teams could use the hybrid to propose novel crystals that are both chemically sensible and structurally sound, widening the search for useful new solids.
Catalog record
| CREATOR | Subhojyoti Khastagir, Kishalay Das, Pawan Goyal, Seung-Cheol Lee, Satadeep Bhattacharjee, Niloy Ganguly |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 84 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2510-23040
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
LLM Meets Diffusion: A Hybrid Framework for Crystal Material Generation Subhojyoti Khastagir, Kishalay Das, Pawan Goyal, Seung-Cheol Lee, Satadeep Bhattacharjee, Niloy Ganguly 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 denoising-based 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 0 0 footnotetext: †Correspondence to Kishalay: [email protected] 1 Introduction Discovery of novel three-dimensional crystal materials with desired chemical properties remains a fundamental challenge in the field of materials design. These materials play a critical role in driving innovations such as development of batteries, solar cells, and semiconductors butler2018machine ; desiraju2002cryptic . Generating crystal materials presents a significant challenge due to its intrinsic complex structure. Unlike molecules, which are typically represented as regular graphs, crystal materials are typically modeled by a minimal unit cell containing all the constituent atoms in different coordinates, repeated infinite times in 3D space on a regular lattice, which makes material structures periodic in nature schutt2014represent . Hence, generating stable crystal material requires the simultaneous prediction of both discrete (atomic types) and continuous components (atomic coordinates and lattice structures). Traditionally, material discovery has relied on either computationally expensive density functional theory (DFT) simulations Jones1989 ; Parr1995 ; Sholl2009 or labor-intensive experimental procedures. However, recent advances in generative modeling have opened up promising directions for designing novel crystal structures with greater efficiency and scalability. Current generative approaches for material generation can be broadly classified into two main categories: (1) autoregressive large language models (LLMs) ( gruver2024fine ) , and (2) denoising-based frameworks, including diffusion models ( xie2021crystal ; luo2023towards ; jiao2023crystal ; zeni2023mattergen ; jiao2024space ; yang2023scalable ) and flow matching techniques ( miller2024flowmm ; sriram2024flowllm ) . Each of these generative models has distinct advantages and drawbacks. LLMs excel at capturing discrete information, making them particularly effective at predicting atomic types and achieving high compositional validity. However, they often face challenges in accurately generating continuous data such as atomic positions and lattice parameters due to limitations in finite precision encoding, leading to reduced structural validity. In contrast, denoising-based models are inherently better at handling continuous variables and preserving geometric equivariances, resulting in higher structural validity. Nevertheless, these models struggle with discrete components, such as correctly identifying atomic types, which leads to lower compositional validity (Table- 1 ). In this work, we aim to harness the complementary strengths of both LLMs and denoising models to mitigate their individual limitations. To this end, we introduce CrysLLMGen, a novel hybrid framework for periodic material generation, which consists of two modules: a fine-tuned LLM and a pre-trained Diffusion Model. Our sampling process begins with the LLM that generates intermediate predictions for atom types, coordinates, and lattice. Given that LLMs excel at modeling discrete information, the generated atom types are well-aligned with the true atomic distributions observed in material datasets. Hence, we retain atomic types and then pass the predicted atomic coordinates and lattice structures to a pre-trained equivatiant diffusion model, which refines and adjusts these continuous components to produce more stable and structurally valid crystal materials. Our simple yet effective hybrid approach enables the joint modeling of multimodal features, both discrete and continuous, of crystal materials, significantly enhancing the validity, stability, and novelty of the generated structures. A major advantage of incorporating large language models (LLMs) is their ability to process natural language prompts, allowing for conditional generation. Additionally, denoising models handle crystal symmetries and the equivariant nature of material data distributions. Importantly, our proposed framework is architecture-agnostic, meaning it can readily accommodate future advancements in both LLMs and denoising networks without requiring substantial modifications. Extensive experiments on benchmark datasets demonstrate that hybrid models can effectively capture both the discrete and continuous aspects of crystal structures, resulting in improvements in both compositional and structural validity. Furthermore, our results show that CrysLLMGen generates 32% and 68% more stable materials compared to state-of-the-art LLM-based models and best-performing denoising model. To sum up, our novel contributions in this work are as follows: • To the best of our knowledge, CrysLLMGen is the first hybrid model that combines large language models (LLMs) with diffusion models for crystal material generation. • Extensive experiments on benchmark datasets demonstrate that CrysLLMGen consistently outperforms state-of-the-art models in both structural and compositional validity. It also generates more stable, unique, and novel crystal structures compared to existing approaches. • CrysLLMGen shows strong generative capability under conditional prompts, effectively producing materials aligning with specified atomic compositions and space group constraints. Validity Metric Denoising Models LLMs LLM+Diffusion CDVAE ( xie2021crystal ) DiffCSP ( jiao2023crystal ) FlowMM ( miller2024flowmm ) LLaMA-2(7B) ( gruver2024fine ) CrysLLMGen(7B) Structural Validity 100 100 96.85 96.40 99.94 Compositional Validity 86.70 83.25 83.19 93.30 93.55 Table 1: Validity Evaluation of Crystal Generation across model classes: Denoising models tend to achieve higher structural validity, while LLMs excel in compositional validity. CrysLLMGen demonstrates a balanced performance, delivering both high structural and compositional validity. 2 Related Work: Crystal Material Generation Earlier works on periodic material generation primarily focused on atomic composition, often overlooking 3D structural details. With the rise of generative models, approaches using VAEs or GANs have aimed to generate 3D periodic structures by representing materials as voxel images court20203 ; hoffmann2019data ; long2021constrained ; noh2019inverse or embedding vectors kim2020generative ; ren2020inverse ; zhao2021high ; however, these methods neither ensure structural stability nor maintain invariance to Euclidean or periodic transformations. Diffusion Models. Recently, equivariant diffusion models have become the leading approach for stable crystal material generation due to their ability to utilize the physical symmetries of periodic structures. Models like CDVAE xie2021crystal and SyMat luo2023towards combine VAEs with score-based denoising networks operating on atomic coordinates through equivariant GNNs, ensuring Euclidean and periodic invariance. Subsequent models, including DiffCSP jiao2023crystal and MatterGen zeni2023mattergen , jointly learn atomic composition, coordinates, and lattice parameters via diffusion frameworks. UniMat yang2023scalable extends this idea with a unified 4D tensor representation that models discrete atom types and continuous coordinates using a probabilistic diffusion model. More recently, text-guided diffusion models such as TGDMat das2025periodic and Chemeleon park2025exploration integrate contextual embeddings from pretrained models like MatSciBERT or CLIP into GNN-based denoising networks to generate stable periodic materials aligned with textual descriptions. Symmetry-Aware Generation. While conventional diffusion models learn