From 4D-STEM data to interpretable physics — an unsupervised learning approach to the charge order physics in TaS₂

The increasing volume and complexity of data from modern probes call for new data-centric approaches for connecting the data to scientific insight. 4-dimensional scanning transmission electron microscopy (4D-STEM) data provides complete imaging of the 2D electron diffraction patterns at each spatial pixel, opening access to rich physics of large scale spatial variation in charge distribution and atomic structure. However, the large dimensionality of the dataset, especially when taken under a variation in a control knob such as temperature, quickly overwhelms traditional mode of data analysis. To harness the new information 4D-STEM can offer, we adopt an unsupervised machine-learning technique, X-ray TEmperature series Clustering (X-TEC) recently developed for voluminous X-ray data [1]. We focus on rich charge density wave ordering phenomenology of transition metal dichalcogenide, specifically, 1T-TaS₂, to distinguish between the commensurate and nearly-commensurate charge density wave. Extending X-TEC to 4D-STEM, we reproduce known charge density wave ordering temperature and find nearly commensurate regions. I will discuss the implications of the findings for the physics of TaS₂ as well as the prospect of using the new approach more broadly.