[Paper Review] End-to-End Crystal Structure Prediction from Powder X-Ray Diffraction

Qingsi Lai, Xu, Fanjie+9 more|arXiv (Cornell University)|Jan 8, 2024

X-ray Diffraction in Crystallography6 citations

TL;DR

XtalNet is an end-to-end deep learning framework that retrieves and conditionally generates crystal structures directly from PXRD patterns, achieving high top-10 performance on a challenging MOF-derived dataset.

ABSTRACT

Powder X-ray diffraction (PXRD) is a prevalent technique in materials characterization. While the analysis of PXRD often requires extensive human manual intervention, and most automated method only achieved at coarse-grained level. The more difficult and important task of fine-grained crystal structure prediction from PXRD remains unaddressed. This study introduces XtalNet, the first equivariant deep generative model for end-to-end crystal structure prediction from PXRD. Unlike previous crystal structure prediction methods that rely solely on composition, XtalNet leverages PXRD as an additional condition, eliminating ambiguity and enabling the generation of complex organic structures with up to 400 atoms in the unit cell. XtalNet comprises two modules: a Contrastive PXRD-Crystal Pretraining (CPCP) module that aligns PXRD space with crystal structure space, and a Conditional Crystal Structure Generation (CCSG) module that generates candidate crystal structures conditioned on PXRD patterns. Evaluation on two MOF datasets (hMOF-100 and hMOF-400) demonstrates XtalNet's effectiveness. XtalNet achieves a top-10 Match Rate of 90.2% and 79% for hMOF-100 and hMOF-400 in conditional crystal structure prediction task, respectively. XtalNet enables the direct prediction of crystal structures from experimental measurements, eliminating the need for manual intervention and external databases. This opens up new possibilities for automated crystal structure determination and the accelerated discovery of novel materials.

Motivation & Objective

• Motivate ab initio crystal structure prediction directly from PXRD without external databases or manual refinement.
• Develop a dual-task framework that can both retrieve similar crystal structures and generate conditioned crystal structures from PXRD data.
• Demonstrate effectiveness on realistic, complex materials (≤100 atoms per unit cell).
• Provide a data-driven pathway for rapid, automated PXRD-based structure determination.

Proposed method

• Two-task architecture: a Contrastive PXRD-crystal Pretraining (CPCP) module for database retrieval via contrastive learning between PXRD features and crystal structures; a Conditional Crystal Structure Generation (CCSG) module based on diffusion models guided by PXRD features.
• PXRD Feature Extractor based on a transformer that encodes peak-focused PXRD data into a PXRD embedding.
• Crystal Structure Network (Modified Equivariant Graph Neural Network) that processes lattice and fractional coordinates and is used for diffusion-based generation.
• Contrastive learning with InfoNCE loss to align PXRD and crystal structure embeddings.
• Diffusion-based generation (DiffCSP framework) for lattice and fractional coordinates conditioned on PXRD embeddings; denoising targets predict lattice and fractional coordinate residuals.
• Training on hMOF-100 with simulated PXRD from MOFs, using top-k hit rate and RMSE as evaluation metrics.

Experimental results

Research questions

• RQ1 Can PXRD patterns be used to retrieve the corresponding crystal structures without external databases?
• RQ2 Can a diffusion-based model generate accurate crystal structures conditioned on PXRD patterns, including complex materials with up to 100 atoms per unit cell?
• RQ3 How well can a joint contrastive and conditional generation framework perform on realistic MOF-like data?
• RQ4 What is the quality of generated structures when ranked by PXRD similarity to ground truth?

Key findings

• Top-10 hit ratio for database retrieval reaches 96.3% on 9,081 test cases.
• Ranked structure generation achieves 95% top-10 match rate with RMSE 0.3985 on the top-10 candidates."
• Ablation studies show that proper PXRD feature fusion and pretraining of the PXRD encoder are critical for performance.
• The framework demonstrates effective generation of metal-connecting motifs and broadly plausible ligand structures, though fine details may diverge from ground truth.