Active-Space Optimization via Reinforcement Learning for Multireference Excited-State Calculations

Roldao et al. (2022), Kabir et al. (2024), and others stress that the choice of active space is crucial for accurate multireference excited-state calculations, but current selection approaches are semi-empirical and system-dependent. This project introduces reinforcement learning (RL) to iteratively construct and refine the active space based on feedback from spectral predictions, energy gaps, and comparison to experimental benchmarks (where available). The RL agent would explore different orbital combinations, learning which choices yield the most reliable results across a range of molecules—π-conjugated systems, transition metal complexes, or biological chromophores. This synthesizes computational chemistry with modern AI, paving the way for “self-optimizing” electronic structure protocols. The impact is in democratizing high-level excited-state calculations, making them more accessible and robust even for non-expert users or for high-throughput screening.

References:

1. Quantum-chemistry study of the ground and excited state absorption of distyrylbenzene: Multi vs single reference methods.. J. C. Roldao, E. Oliveira, Begoña Milián‐Medina, J. Gierschner, D. Roca‐Sanjuán (2022). Journal of Chemical Physics.
2. Electronic Structure Methods for Simulating Flavin’s Spectroscopy and Photophysics: Comparison of Multi-reference, TD-DFT, and Single-Reference Wave Function Methods. M. Kabir, Paulami Ghosh, S. Gozem (2024). Journal of Physical Chemistry B.