Scholarlink

<p>Artificial intelligence (AI) methods are increasingly being applied to X-ray diffraction (XRD) for crystal phase identification. However, in multiphase materials, peak overlap and noise masking of weak diffraction signals make it challenging for machine learning, particularly deep learning approaches, to accurately distinguish different phases. To address this issue, a dual-branch framework termed the wavelet-enhanced non-pooling convolutional neural network (WNPCNN) is proposed in this work. In the first branch, discrete wavelet transform (DWT) and inverse wavelet transform (IWT) are used to denoise and reconstruct diffraction patterns, which are then stacked with the original diffraction patterns and processed by a non-pooling CNN to preserve relative peak positions and to enhance sensitivity to weak diffraction signals. The second branch transforms diffraction patterns into real-space pair distribution functions (PDFs) via Fourier-transform (FT), enabling complementary feature extraction of interatomic correlations through convolution and pooling operations. Phase predictions from both branches are fused by a confidence-weighted strategy. The results show that WNPCNN significantly outperforms conventional CNNs, achieving higher accuracy on both simulated and experimental datasets, especially under severe peak overlap and imbalanced phase ratios. These findings demonstrate the potential of WNPCNN as a scalable framework for robust, high-throughput multiphase XRD phase identification in materials discovery.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A Study on Phase Prediction Methods Based on Wavelet Transform and Deep Learning by XRD Data

  • 期刊论文

  • 发表日期：

    2026年3月27日

查看全文
  • Yan Xu,
  • Shangshang Wu,
  • Baihan Lv,
  • Yanyan Zhu,
  • Huiran Zhang,
  • Shengzhou Li,
  • Dongbo Dai

摘要

Artificial intelligence (AI) methods are increasingly being applied to X-ray diffraction (XRD) for crystal phase identification. However, in multiphase materials, peak overlap and noise masking of weak diffraction signals make it challenging for machine learning, particularly deep learning approaches, to accurately distinguish different phases. To address this issue, a dual-branch framework termed the wavelet-enhanced non-pooling convolutional neural network (WNPCNN) is proposed in this work. In the first branch, discrete wavelet transform (DWT) and inverse wavelet transform (IWT) are used to denoise and reconstruct diffraction patterns, which are then stacked with the original diffraction patterns and processed by a non-pooling CNN to preserve relative peak positions and to enhance sensitivity to weak diffraction signals. The second branch transforms diffraction patterns into real-space pair distribution functions (PDFs) via Fourier-transform (FT), enabling complementary feature extraction of interatomic correlations through convolution and pooling operations. Phase predictions from both branches are fused by a confidence-weighted strategy. The results show that WNPCNN significantly outperforms conventional CNNs, achieving higher accuracy on both simulated and experimental datasets, especially under severe peak overlap and imbalanced phase ratios. These findings demonstrate the potential of WNPCNN as a scalable framework for robust, high-throughput multiphase XRD phase identification in materials discovery.