Topological Machine Learning Unveils Hidden Reaction Pathways in Nanocrystal Synthesis

Abstract

Uncovering reaction pathways from analytical data such as UV-vis spectra remains a central challenge in nanocrystal synthesis, where transient and ill-defined intermediates complicate mechanistic analysis. Conventional approaches, reliant on manual spectral feature extraction and expert interpretation, are prone to bias and often overlook critical events. Here we present a machine learning framework that integrates transformer-based data augmentation with topological manifold learning to objectively elucidate reaction pathways directly from raw, high-dimensional spectroscopic data. The key insight is that the topology of the data manifold reflects the structure of the underlying reaction pathway. Applied to ex-situ UV-vis data sets of indium arsenide nanocrystal synthesis, this approach reconstructs the complete reaction landscape, identifying previously unreported metastable intermediates and revealing how chemical additives modulate intermediate formation to steer pathway selection. Broadly adaptable to diverse analytical data, this topological learning framework provides a generalizable strategy for mechanistic discovery and predictive control in complex chemical systems.