Harnessing interpretable and unsupervised machine learning to address big data from modern X-ray diffraction.

Area of Science:

• Materials Science
• Condensed Matter Physics
• Data Science

Background:

• Crystalline materials contain vast information, especially concerning collective electronic behavior and fluctuations.
• Advances in X-ray sources and detectors enable capturing more of this information.
• Analyzing large datasets from X-ray diffraction (XRD) experiments presents a significant challenge for human researchers.

Purpose of the Study:

• To develop an unsupervised machine learning approach for analyzing large XRD datasets.
• To automatically extract charge density wave order parameters and detect intraunit cell ordering and fluctuations.
• To demonstrate the application of this method in uncovering new scientific principles from experimental data.

Main Methods:

• Developed X-ray diffraction (XRD) temperature clustering (X-TEC), an unsupervised machine learning algorithm.
• Applied X-TEC to analyze high-volume XRD measurements taken at multiple temperatures.
• Benchmarked X-TEC using data from (Ca,Sr)3Rh4Sn13 and applied it to Cd2Re2O7.

Main Results:

• X-TEC successfully extracted charge density wave order parameters and detected atomic ordering and fluctuations.
• Analysis of Cd2Re2O7 revealed insights into its structural phase transitions and associated Goldstone mode.
• Identified specific atomic displacements in Cd2Re2O7, suggesting an electronic origin for structural order.

Conclusions:

• The X-TEC approach can automatically extract valuable scientific information from large XRD datasets.
• This method facilitates in operando data analysis and dynamic experimental refinement.
• Unprecedented atomic-scale knowledge can be gained by integrating X-TEC results with physical principles.