Fig. 1
History of photoemission-data quantity and schematics of causal discovery. (a) Schematics on the history of data types (top) and data capacity (bottom) in photoemission spectroscopy. In the 1980s, photoemission data were mainly recorded as a 1D spectrum, I(E). In the 2000s, due to the development of advanced electron analyzers that enable the simultaneous angular collection of photoelectrons, photoemission data were obtained as a 2D image, I(E, k1). Further, acquiring a series of 2D images enabled the construction of 3D chunk data I(E, k1, k2). By the 2020s, due to an advancement of highly brilliant micro/nano-spot light sources together with the development of state-of-the-art spin detectors, photoemission data contain multi-dimensional (more than 4D) elements, including lateral position (x, y) on the sample surface and spin component (Sx, Sy, Sz). Consequently, data capacity increased from Kbytes (in the 1980s) to Gbytes/Tbytes (in the 2020s). (b) Schematic of a conventional data analysis flow. Researchers or artificial intelligence (AI)-based approaches extract correlations between input variables, from which researchers try to estimate causality using regression models or by relying on their intuition and experience. (c) Schematic of AI-based causal discovery from the data. AI-based methods identify causal relationships, i.e., cause variables and their corresponding effect variables with/without predefined conditions, and generate causal graphs without expert intuition or experience. Researchers can notice underlying scientific laws from these causal graphs.
