[Paper Review] XERUS: An open-source tool for quick XRD phase identification and refinement automation

Pedro Baptista de Castro, Kensei Terashima+3 more|arXiv (Cornell University)|Dec 9, 2021

Machine Learning in Materials ScienceMaterials Science43 references12 citations

TL;DR

XERUS is an open-source Python tool for rapid, automated XRD phase identification and refinement that leverages open crystallographic databases (e.g., MP, OQMD, COD) to enable on-the-fly structure retrieval and pattern simulation. It combines similarity-based phase matching with iterative Rietveld refinement and peak subtraction, achieving high accuracy in multi-phase identification without pre-trained models, as validated on diverse experimental datasets and a Li-Mn-O-F mixture benchmark.

ABSTRACT

Analysis of XRD diffraction patterns is one of the keystones of materials science and materials research. With the advancement of data-driven methods for materials design, candidate materials can be quickly screened for the study of a desired physical property. Efficient methods to automatically analyze and identify phases present in a given pattern, are paramount for the success of this new paradigm. To aid this process, the open source python package Xray Estimation and Refinement Using Similarity (XERUS) for semi-automatic/automatic phase identification is presented. XERUS takes advantages of open crystal structure databases, not relying on proprietary databases, to obtain crystal structures on the fly, being then chemical space agnostic. By wrapping around GSASII, it can automatically simulate patterns and calculate similarity measures used for phase identification. Our approach is simple and quick but also applicable to multiphase identification, by coupling the similarity calculations with quick refinements followed by an iterative peak removal process. XERUS is shown in action in four different experimental datasets and also it is benchmarked against a recently proposed deep learning method for a mixture dataset covering the Li-Mn-O-F chemical space. XERUS will be freely available on https://www.github.com/pedrobcst/Xerus/

Motivation & Objective

• Address the bottleneck in materials research caused by time-consuming manual XRD analysis in high-throughput materials screening.
• Overcome limitations of deep learning-based XRD methods that rely on proprietary databases (e.g., ICSD) and require extensive retraining for new chemical spaces.
• Develop a flexible, chemical-space-agnostic tool that dynamically retrieves crystal structures from open databases to support real-time phase identification.
• Enable semi-automatic and one-shot Rietveld refinement for rapid synthesis validation in data-driven materials discovery workflows.
• Provide an extensible, open-source framework compatible with Jupyter notebooks and experimental pipelines for interactive or automated use.

Proposed method

• Uses open-access crystal structure databases (Materials Project, OQMD, AFLOW, COD) via APIs to retrieve structures on-demand for any given chemical composition.
• Employs pymatgen to parse and cache crystal structures locally in a MongoDB database for efficient reuse.
• Performs pattern simulation using GSAS-II for candidate structures based on input XRD data and elemental composition.
• Calculates pattern similarity (e.g., correlation or Rwp) between experimental and simulated patterns to rank candidate phases.
• Applies iterative refinement and peak removal: identifies the highest-correlated phase, refines it, subtracts its peaks, and repeats for multi-phase systems.
• Integrates with GSAS-II for Rietveld refinement and provides visualization and optimization interfaces via Jupyter notebooks.

Experimental results

Research questions

• RQ1 Can a similarity-based, refinement-driven approach achieve accurate XRD phase identification without relying on pre-trained deep learning models?
• RQ2 How effective is an open-source, database-agnostic method for phase identification across diverse and unknown chemical spaces?
• RQ3 Can iterative peak subtraction and refinement improve accuracy in multi-phase XRD analysis compared to single-step similarity matching?
• RQ4 How does XERUS perform on real experimental datasets and benchmark datasets like the Li-Mn-O-F mixture compared to state-of-the-art deep learning methods?
• RQ5 To what extent can XERUS support rapid, one-shot Rietveld analysis in high-throughput materials research workflows?

Key findings

• XERUS achieved 92% accuracy in multi-phase classification on a Li-Mn-O-F mixture dataset, matching the performance of a deep learning model that required 20 hours of computational time for training and simulation.
• For a HoB2.1 alloy, XERUS correctly identified HoB2 and HoO as the main phases with Rwp values of 10.46% and 10.54%, respectively, and confirmed the presence of a minor phase (Ho2O3) with 6.7% weight fraction.
• In a La2−xSrxCuO4 sample (x=0.25), XERUS identified the primary phase (La2CuO4) with Rwp = 18.16% and correctly assigned a secondary phase (Sr0.3La3.7Cu2O7.2) with 6.7% weight fraction.
• For a Mo-Ho-C alloy with target composition HoMoC2, XERUS correctly identified HoMoC2 as the top candidate (Rwp = 15.17%) and provided accurate phase quantification across multiple runs.
• XERUS demonstrated robust performance on complex mixtures, such as LiMn2O4 + LiF + Li2MnO3, where it correctly identified the dominant phases with Rwp values below 10% and provided consistent weight percent estimates.
• The tool successfully handled challenging cases with overlapping peaks and non-stoichiometric phases, showing high stability and reproducibility across multiple runs (n_runs = 2–3) and parameter settings (δ = 1.3).