Rethinking Crystal Symmetry Prediction: A Decoupled Perspective

TL;DR

The paper tackles the SPC-driven challenge of predicting crystal space groups from PXRD data. It proposes XRDecoupler, a decoupled framework that injects chemical knowledge via multidimensional superclasses (crystal systems, Bravais lattices, and point groups) and combines a hierarchical PXRD pattern learner with a multi-objective optimization to balance learning across subsymbolic properties. The approach comprises two core components: Superclass-Guided Optimization, which maximizes discriminative information across sub-properties using a Pareto-optimized gradient direction, and Hierarchical PXRD Pattern Learning, which fuses local peak relations and global pattern context into a joint representation $E=\text{Concat}(E_{global},E_{local})$. Empirical results on MOF, CoREMOF, and InorganicData show XRDecoupler achieving state-of-the-art accuracy and better generalization to out-of-domain data, validating the value of embedding chemical principles into symmetry prediction and of balancing multiple superclass objectives.

Abstract

Efficiently and accurately determining the symmetry is a crucial step in the structural analysis of crystalline materials. Existing methods usually mindlessly apply deep learning models while ignoring the underlying chemical rules. More importantly, experiments show that they face a serious sub-property confusion SPC problem. To address the above challenges, from a decoupled perspective, we introduce the XRDecoupler framework, a problem-solving arsenal specifically designed to tackle the SPC problem. Imitating the thinking process of chemists, we innovatively incorporate multidimensional crystal symmetry information as superclass guidance to ensure that the model's prediction process aligns with chemical intuition. We further design a hierarchical PXRD pattern learning model and a multi-objective optimization approach to achieve high-quality representation and balanced optimization. Comprehensive evaluations on three mainstream databases (e.g., CCDC, CoREMOF, and InorganicData) demonstrate that XRDecoupler excels in performance, interpretability, and generalization.

Figures (12)

• Figure 1: The SPC problem in the space group identification. Different colored blocks represent various symmetry classification systems, such as lattice types and point group types. We illustrate four space groups that current methods often confuse: I4, I$\overline{\texttt{4}}$, P4, and P$\overline{\texttt{4}}$. These space groups are intertwined and may belong to a coarser classification. We also present four representative crystal samples, i.e., PELQUU, OVOSOJ, TUTBUG01, and YIWTIK, demonstrating the capability of our method to decouple these confusions.
• Figure 2: Accuracy statistics of samples with misidentified space groups on three sub-properties (e.g., lattice type, crystal system, and point group) on the SOTA method (XRDMamba yu2024xrdmamba) and our proposed XRDecoupler.
• Figure 3: Overview of XRDecoupler.
• Figure 4: Evaluation on MOF subset (left) and MOF-Balanced subset (right) of CCDC dataset with SOTA methods. Bold indicates the best performance while underline indicates the second best. ($+$) and ($-$) indicate the the relative gain with CNN.
• Figure 5: Trend of the model's training loss. (left) The conventional optimization process of the model on the space group and superclass. (middle) The optimization process of the model on the space group and superclass after introducing the gradient-based optimization method. (right) T-SNE Visualization Analysis of XRDecoupler in the training set.

Theorems & Definitions (12)

• Proposition 1
• Proposition 2
• Definition 1: Difference $Diff$ in mutual information
• Proposition 3
• Definition 2: Solution $W$
• Definition 3: Solution $W^*$
• Proposition 2: Proposed in paper
• proof
• Definition 1: Proposed in paper
• proof