Abstract:
X-ray absorption spectroscopy is a premier element-specific experimental technique for materials characterization. Specifically, X-ray absorption near edge structure (XANES) encodes rich local structural and chemical information around absorber sites, making it a powerful tool for probing physical and chemical processes in different domain fields. However, the correlation between XANES spectral features and the underlying local structural motifs or electronic descriptors is obscure. Recent progress in materials discovery using smart automation and in situ / operando experiments underscores emerging challenges and opportunities of spectral analysis in real time. This challenge cannot be tackled by empirical fingerprints, which have practical limitations on data availability and diversity, or first-principles simulations alone due to the high computational cost. Here I discuss the recent progress of the data-driven spectral analysis pipeline by combining first-principles theory, software development, high throughput simulation, and data analytics tools to decipher the structure-spectrum relationship. I highlight the recent work where graph neural network machine learning models are developed to predict 3d transition metal K-edge XANES of materials from their atomic structures. We find that the best performance can be achieved by training a universal model on all the elements and subsequently fine-tuning on element- specific data. We also find this transfer learning strategy can be applied across different fidelities of methods, which offers a generalizable framework in deep-learning models for materials science.

Virtual Link:
Microsoft Teams Need help?
Join the meeting now
Meeting ID: 262 477 287 747
Passcode: BFsHUu
Dial in by phone
+1 630-556-7958,,189052388# United States, Big Rock
Find a local number
Phone conference ID: 189 052 388#