CrystalX: Ultra-Precision Crystal Structure Resolution and Error Correction Using Deep Learning

TL;DR

This deep learning model revolutionizes the time frame for crystal structure analysis, slashing it down to seconds, and marks the beginning of a new era in automating routine structural analysis within self-driving laboratories.

Abstract

Atomic structure analysis of crystalline materials is a paramount endeavor in both chemical and material sciences. This sophisticated technique necessitates not only a solid foundation in crystallography but also a profound comprehension of the intricacies of the accompanying software, posing a significant challenge in meeting the rigorous daily demands. For the first time, we confront this challenge head-on by harnessing the power of deep learning for ultra-precise structural analysis at the full-atom level. To validate the performance of the model, named CrystalX, we employed a vast dataset comprising over 50,000 X-ray diffraction measurements derived from authentic experiments, demonstrating performance that is commensurate with human experts and adept at deciphering intricate geometric patterns. Remarkably, CrystalX revealed that even peer-reviewed publications can harbor errors that are stealthy to human scrutiny, yet CrystalX adeptly rectifies them. This deep learning model revolutionizes the time frame for crystal structure analysis, slashing it down to seconds. It has already been successfully applied in the structure analysis of newly discovered compounds in the latest research without human intervention. Overall, CrystalX marks the beginning of a new era in automating routine structural analysis within self-driving laboratories.

Figures (4)

• Figure 1: The CrystalX neural network approach to crystal structure analysis. The inputs consist of coarse electron density peaks, which are processed using an Equivariant Transformer to extract geometric interaction patterns for heavy-atom (non-hydrogen atoms) determination. Following this, a separate Equivariant Transformer is developed to capture interactions for hydrogen determination. The dataset is derived from automatically phased coarse electron density in experimental diffraction measurements, with expert human interpretations serving as annotations.
• Figure 2: Performance overview on large-scale real experimental data.(a) Element-level F1 score for non-hydrogen elements. (b) Unit-cell-level accuracy for non-hydrogen elements. (c) F1 score for prevalent hydrogen atoms. (d, e) Comparison of crystallographic metrics (i.e., $R_1$ and $S$) between the model and human experts. (f, g) Complex structure analysis (with up to 370 heavy atoms) automated by the model.
• Figure 3: Visualization and interpretability of the model's underlying behavior. (a-d) Visualization of attention maps from Equivalent Transformers using Attention Rollout attn-rollout with class activation demonstrates precise identification of elements like carbon, nitrogen, oxygen, and metals (b,d). This identification is achieved solely by analyzing the geometric patterns of coarse electron density, without the need for any prior information. The interaction intensity is shaded according to class activation levels. (e,f) Two-dimensional t-SNE t-sne projections of the model's learned representations for non-hydrogen atom identification (e) and hydrogen atom identification (f).
• Figure 4: CrystalX corrects errors in JCR Q1 journals.a, Error corrections through crystallographic comparison of the structure analysis results between the model and human experts. b-d, The model corrects three types of errors found in the published literature: (b) misidentification of atoms with similar charges, (c) incorrect placement of hydrogen atoms, and (d) missing hydrogen atoms. These errors are often difficult to detect, passing the rigorous reviews of JCR Q1 journals. The model's analysis offers both improved structural rationality and superior crystallographic metrics.