Accurate Prediction of ECD Spectra of Silver-Based Chiral Complexes Using Femtosecond-Informed Machine Learning from...

Electronic circular dichroism (ECD) stands out as a highly sensitive spectral probe for chiral systems, facilitating molecular configuration determination, analysis of metal-ligand interactions, and the design of chiral optical materials. However, traditional machine learning struggles to predict ECD spectra of metal complexes accurately because multiple coupled electronic transitions, complex local-environment perturbations, limited data availability, and model interpretability constraints introduce significant uncertainty. To address these issues, we propose a physics-inspired machine learning framework that integrates "femtosecond charge dynamics - ECD spectra - electronic transitions." This approach enables predictions of chiral spectra for silver-based metal complexes with quantitatively close to the corresponding TDDFT references for the present data set but at just 1/470th of the computational time. This framework serves as a interpretable and computationally efficient framework within the present silver-based chemical space for tasks such as absolute configuration determination, chiral catalysis screening, and reverse design of optical materials. It achieves precise mapping between real-time femtosecond charge dynamics and the ECD spectra of chiral silver clusters and Ag-DNA complexes. A systematic evaluation using a self-constructed data set reveals that our MC-HG model increases computational speed by approximately 470 times compared to traditional TDDFT methods. Furthermore, the PSF-ECD model enhances spectral shape similarity by a factor of 3 and reduces error indices by 45% compared to single traditional machine learning models. By integrating the transition contribution map (TCM) with Mulliken charge flow visualization, we establish an interpretable link between learned spectral shapes and specific orbital transitions and charge transfer channels. This offers an efficient pathway for evaluating the chiral optical response of metal complexes without the need for repeated real-time TDDFT calculations.