Vision Transformers for X-Ray Spectroscopy and Fluctuation Analysis*

The combination of self-supervised pre-training on broad data at scale, fine-tuning via supervised and reinforcement learning, and the advent of large, highly-parallelized transformer architectures has led to a proliferation of so-called foundation models and their application in nearly every field of machine learning over the last three years. In the broad field of computer vision, vision transformers (ViT) have quickly become the state of the art in image classification and generative modeling. Yet, the ubiquitous task of image-segmentation, pattern recognition, and transduction in x-ray spectroscopy, as well as experimental physics more broadly, remains an unexplored domain for ViT. X-ray photon fluctuation spectroscopy (XPFS) is an advanced x-ray spectroscopy technique which allows for the direct measurement of fluctuations in a material on a nanosecond timescale. In this work we present a generative technique using ViT which can instantaneously convert detector images to photon maps in the regime of low-photon count spectroscopy for XPFS, showing this model outperforms all current state of the art (classical and ML) techniques, and can also be fine-tuned during an experiment to adapt to new detector parameters. Moreover, we show that a foundation model fine-tuned on x-ray spectroscopy data is able to directly predict the intermediate scattering function of a material given a series of raw detector images from an XPFS experiment, a task which has never been accomplished by any existing technique.