Rigaku logo
XRD
Read Article

AI Analysis Basic Course - First Installment: Neural Network Application to Phase Identification in Powder X-ray Diffraction

Toshihide Shibasaki, Takumi Ohta, and Akihiro Himeda

Summer 2025 Volume 41, No. 2 , 06-10

In recent years, there have been significant improvements in AI technology, especially in neural networks. We describe profile-based phase identification using neural networks, which does not require peak search. Using cements and excipients as examples, we report that neural networks can be used to identify crystalline phases more accurately even when analysis by the conventional method is difficult.

Highlights

  • Neural-network phase identification works directly from the full XRD profile, so it does not depend on peak search accuracy.
  • This profile-based approach performed better than conventional search/match for difficult materials such as cement and pharmaceutical excipients.
  • For cement test data, the neural-network method achieved a much higher F1 score overall, and it was especially better at detecting minor phases.
  • For excipient test data, the biggest improvement was in identifying micro-crystalline components, which are often hard to resolve because of broad peaks and poor crystallinity.
  • Amorphous-phase identification remained challenging even with neural networks, showing that background treatment is still a major limitation.

Summary

Phase identification is one of the most common uses of powder XRD, but it becomes difficult when peaks overlap, broaden, or are hard to separate clearly. Traditional search/match methods depend on finding peaks first and then comparing those peaks with a database. That works well in many cases, but it can miss the correct phases when the diffraction pattern is messy or poorly crystalline.

A neural-network approach can improve this by using the whole diffraction profile instead of relying only on a peak list. In this work, a vision transformer model was used to examine XRD patterns directly and estimate which phases were present. Training data were created from simulated diffraction profiles, including mixtures and variations in factors such as lattice constants, crystallite size, and preferred orientation, so the model could learn realistic pattern changes.

When tested on cement and pharmaceutical excipient data, the neural-network method produced more accurate phase identification than conventional search/match. It was especially strong for materials with overlapping peaks or poor crystallinity, although amorphous components were still difficult to identify reliably. Overall, the results show that AI-based profile analysis can make phase identification more robust in cases where conventional peak-based methods struggle.

Frequently asked questions

The main problem is that conventional phase identification often starts with peak search. When peaks are heavily overlapped, broadened, or weakened by poor crystallinity, the peak list can be inaccurate. That can lead to missed phases or wrong candidates. A neural-network approach avoids that dependency by analyzing the diffraction profile more directly.

Conventional search/match converts the diffraction pattern into peak positions and intensities, then compares that list with a database. The neural-network method takes the measured profile, or a preprocessed version of it, as the input. Because it uses the profile itself rather than only extracted peaks, it is less sensitive to peak-search errors.

A vision transformer, or ViT, was used. This model is widely known from image recognition. It divides the input into smaller sections and learns relationships between them. Applied to powder XRD, it can recognize phase-related features across the full pattern, including information that may be spread across multiple peaks or broader profile regions.

The training data were not limited to measured patterns. Simulated diffraction profiles were generated using the fundamental parameter method. Single-phase patterns were varied by changing properties such as lattice constants, crystallite size, and preferred orientation, then combined randomly to create mixed-phase patterns. This gave the model a large and varied training set that reflected realistic changes in profile shape.

Cements are difficult because they contain several major components with similar diffraction patterns and multiple polymorphs. That creates heavy peak overlap and makes it hard for conventional search/match to distinguish the right phases. The neural-network method showed better overall phase identification and was especially better for minor phases, which are easier to miss.

Excipients are often poorly crystalline and may contain micro-crystalline and amorphous components. That leads to broad peaks and diffuse features, which are difficult for conventional peak-based analysis. The neural-network method performed especially well for micro-crystals, showing that profile-based analysis is useful when crystallinity is low.

No. Amorphous phases were still difficult to identify reliably. Even though the neural-network results were better than search/match, performance for amorphous components remained limited. A key issue is that background subtraction can remove or distort the broad features that distinguish amorphous material, making those phases harder to separate from the background signal.

It shows that AI can help generate better phase candidates in cases where conventional methods struggle, especially when patterns are complicated by overlap, polymorphism, or poor crystallinity. That can improve follow-up analysis, including quantitative work such as Rietveld refinement, because better phase selection at the start usually leads to better final results.

Related products

Subscribe to the Bridge newsletter

Stay up to date with materials analysis news and upcoming conferences, webinars and podcasts, as well as learning new analytical techniques and applications.

Contact Us

Whether you're interested in getting a quote, want a demo, need technical support, or simply have a question, we're here to help.

Contact Sales
Contact Service
Office Locations