On the Robustness and Transferability of Statistical Structure Sampling for Machine Learning EXAFS

Abstract

Machine learning inversion of extended X-ray absorption fine structure (EXAFS) into structural descriptors relies on data sets generated from ab initio molecular dynamics (AIMD), which are computationally expensive and system-specific. Here we evaluate the Objective Neural Network for EXAFS (ONNE), a synthetic training data framework that statistically samples configurational spaces. We compare ONNE datasets with AIMD data in terms of configurational diversity, representativeness, and computational cost. Models trained on ONNE data achieve coordination number predictions comparable to AIMD-trained models while requiring 99.37% lower computational costs. Fine-tuning the pretrained ONNE model with a small number of AIMD configurations further improves structural predictions and reduces uncertainty while maintaining a lower computational expense. We demonstrate that ONNE can serve as an efficient pretraining strategy for ML-EXAFS, reducing the dependence on expensive simulations and thus informing future ML-EXAFS investigations by clarifying the advantages, limitations, and applicability of ONNE relative to MD approaches.