Dec 3, 2024
8:00pm - 10:00pm
Hynes, Level 1, Hall A
Amalya Johnson1,2,Chris Fajardo2,Leena Sansguiri2,Weike Ye2,Steven Torrisi2
Stanford University1,Toyota Research Institute2
Amalya Johnson1,2,Chris Fajardo2,Leena Sansguiri2,Weike Ye2,Steven Torrisi2
Stanford University1,Toyota Research Institute2
SpectraScope is a toolkit for materials characterizaiton from spectral data using interpretable machine learning models. It is both a python package and a web app, allowing for its easy accessibility by both experimental and computational materials researchers. The software provides a framework for feature generation, feature selection, and model training. It can currently be used with spectra from experiments such as Raman, x-ray diffraction, x-ray absorption, pair distribution functions, infrared spectra, and optical absorption spectra, and will be expanded to work with two-dimensional microscopy data as well. Additionally, SpectraScope can be applied to time-series data. It has been used to predict coordination number and bond length from x-ray absorption spectra of transition metal oxide structures. Through feature selection, SpectraScope identifies regions of spectra that are important for prediction. This helps shed light on the physical relationships between spectra and the characterization. Additionally, SpectraScope can use multiple datatypes at once for prediction, which can help identify the relationships between different spectra and how they may impact model performance. This talk will outline the software's technical details and accessibility.