Decoding Configurational Populations and Temperature of Flexible Chiral Molecules from Chiral Optical Spectra with Machine Learning

Quantitative determination of configurational populations in flexible chiral molecules is challenging due to spectral overlap, conformational averaging, and random noise. Here, we present a multitask residual network that integrates quantum chemical calculations, spectral simulation, and deep learning to achieve configurational inversion from spectroscopic data. As a proof-of-concept study, the neural network uses DFT-computed ultraviolet–visible absorption and electronic circular dichroism spectra as input to simultaneously predict the populations of configurational isomers and the thermodynamic temperature of the system. Demonstrated on the flexible chiral drug molecule labetalol, the method achieves high accuracy: under 10% spectral noise, the root-mean-square error for configurational populations is about 8.2%, and the temperature error is below 3.4 °C. The successful inversion demonstrates that machine learning can effectively bridge spectroscopic observables and statistical thermodynamics, providing a rapid, data-driven paradigm for resolving structural distributions in complex flexible chiral systems.