KAN-enhanced contrastive learning: the accelerator of crystal structure identification from XRD patterns

Accurate crystal structure determination underpins materials discovery, yet powder X-ray diffraction (XRD) analysis still depends on expert-driven, iterative fitting that limits scalability for high-throughput and autonomous experiments. We introduce XRD-Crystal Contrastive Pretraining (XCCP), a physics-guided contrastive learning framework that aligns PXRD patterns with candidate crystal structures in a shared embedding space to enable efficient structure retrieval and symmetry inference. XCCP employs a dual-expert XRD encoder with a Kolmogorov-Arnold Network (KAN) projection head. A low-angle branch captures long-length-scale signatures, while a wide-angle branch encodes dense, symmetry-governed fingerprints. Attribution and perturbation analyses show that the KAN head concentrates evidence on physically meaningful Bragg reflections rather than background-dominated regions, improving robustness to peak-shape variations. We further introduce similarity-based confidence scores to flag potentially unreliable predictions in open-set settings. Without elemental priors, XCCP achieves 46.42% top-1 accuracy for structure retrieval and 60.85% accuracy for space-group identification. When chemical composition is available for elemental pre-screening, performance increases to 88.98% and 93.39%, respectively. XCCP also generalizes to compositionally similar multi-principal element alloys and enables zero-shot transfer to experimental patterns. These results establish XCCP as an interpretable, confidence-aware, and scalable paradigm for XRD analysis, enabling high-throughput screening, rapid candidate shortlisting, and integration with autonomous laboratory workflows.