Toward Class Imbalance and Uncertainty in Powder XRD Analysis: A Dual-Channel Fusion Network for Space Group...

Area of Science:

• Crystallography and Materials Science
• Computational Chemistry
• Machine Learning Applications

Background:

• Accurate space group identification from powder X-ray diffraction (pXRD) is crucial for materials discovery but faces challenges like peak overlap and data scarcity.
• Existing methods struggle with the 230-class classification problem and imbalanced datasets inherent in crystallographic data.

Purpose of the Study:

• To develop a robust and interpretable automated method for space group classification from pXRD data.
• To address data imbalance and scarcity issues in crystallographic datasets.
• To provide uncertainty estimation for model predictions to enhance reliability.

Main Methods:

• A physics-informed data augmentation pipeline was designed to address data scarcity and class imbalance.
• A dual-channel fusion uncertainty-aware network (DFUN) was proposed, integrating convolutional features and peak descriptors.
• A hybrid loss function (Focal Loss + Label Smoothing) and Monte Carlo Dropout for uncertainty estimation were employed.

Main Results:

• DFUN demonstrated superior performance over baseline methods on simulated and public pXRD datasets (opXRD, RRUFF).
• The model achieved accurate space group classification and provided reliable uncertainty estimates.
• The integrated approach effectively handled complex crystallographic data characteristics.

Conclusions:

• DFUN offers a robust and interpretable solution for high-throughput automated crystallographic analysis using pXRD.
• The uncertainty-aware predictions enhance the trustworthiness of automated materials discovery pipelines.
• The proposed methods advance the field of computational crystallography and machine learning in materials science.