Technology Category
Machine Learning: Graph-based Machine LearningReasoning under Uncertainty: Graphical ModelsKnowledge Representation and Reasoning: Geometric, Spatial, and Temporal Reasoning
PDDFormer: Pairwise Distance Distribution Graph Transformer for Crystal Material Property Prediction
🤖 AI Summary
At finite temperatures, atomic thermal vibrations disrupt the static lattice periodicity, rendering fixed-lattice-parameter graph representations sensitive and discontinuous to small atomic displacements. While pairwise distance distributions (PDDs) offer robustness, they neglect elemental identity and incur high computational cost. To address this, we propose element-weighted PDD (WPDD) and unit-cell-level PDD (UPDD), the first formulations that jointly encode elemental information and preserve periodic boundary consistency. We further design a PDD-driven Graph Transformer architecture featuring unit-cell-aware attention, ensuring graph continuity and prediction stability under atomic perturbations while retaining the theoretical completeness of PDDs. Evaluated on multiple materials property prediction benchmarks, our method achieves state-of-the-art performance and significantly enhances robustness against lattice distortions.
Technology Category
Application Category
📝 Abstract
The crystal structure can be simplified as a periodic point set repeating across the entire three-dimensional space along an underlying lattice. Traditionally, methods for representing crystals rely on descriptors like lattice parameters, symmetry, and space groups to characterize the structure. However, in reality, atoms in material always vibrate above absolute zero, causing continuous fluctuations in their positions. This dynamic behavior disrupts the underlying periodicity of the lattice, making crystal graphs based on static lattice parameters and conventional descriptors discontinuous under even slight perturbations. To this end, chemists proposed the Pairwise Distance Distribution (PDD) method, which has been used to distinguish all periodic structures in the world's largest real materials collection, the Cambridge Structural Database. However, achieving the completeness of PDD requires defining a large number of neighboring atoms, resulting in high computational costs. Moreover, it does not account for atomic information, making it challenging to directly apply PDD to crystal material property prediction tasks. To address these challenges, we propose the atom-Weighted Pairwise Distance Distribution (WPDD) and Unit cell Pairwise Distance Distribution (UPDD) for the first time, incorporating them into the construction of multi-edge crystal graphs. Based on this, we further developed WPDDFormer and UPDDFormer, graph transformer architecture constructed using WPDD and UPDD crystal graphs. We demonstrate that this method maintains the continuity and completeness of crystal graphs even under slight perturbations in atomic positions.
Problem
Research questions and friction points this paper is trying to address.
Static crystal representations fail under atomic position fluctuations
Pairwise distance distribution methods have high computational costs
Existing PDD methods lack atomic information for property prediction
Innovation
Methods, ideas, or system contributions that make the work stand out.
Atom-weighted pairwise distance distribution for crystal graphs
Unit cell pairwise distance distribution integration
Matrix-based message passing with reduced computational cost
🔎 Similar Papers
No similar papers found.
