Transfer Analysis of Machine Learning Models Trained on Simulated XRD Patterns to Real Powder XRD Experimental Data

Accurate crystal structure identification is crucial for material property prediction, functional design, and the discovery of new materials. Machine learning models trained on simulated X-ray diffraction (XRD) patterns based on the International Crystal Structure Database (ICSD) have shown promising results under idealized conditions. However, real experimental data often contain complex factors such as peak shifts, peak broadening, multiphase overlap, and background noise, which can lead to performance degradation when applying these models directly. In this study, we constructed a hybrid feature learning model by training on ICSD-based standard XRD patterns, combining handcrafted physical features with deep convolutional features. We then systematically evaluated its transfer performance on real-world powder XRD data collected from various sources. The results showed that, despite domain differences, the model achieved a crystal system classification accuracy of 91.5 %, a space group Top-1 accuracy of 76.8 %, and a Top-5 accuracy of 89.2 %, demonstrating good transferability. Through detailed error analysis, we found that multiphase coexistence, peak loss, and instrumental noise were the main sources of classification errors. Furthermore, we discuss potential strategies to improve transfer performance, including transfer learning, data augmentation, and the integration of physical prior knowledge. This work provides new insights into applying machine learning models trained on standard data for the intelligent analysis of real experimental XRD patterns, thereby promoting the broader application of AI in materials research.