XMCQDPT2-fidelity transfer-learning potentials and a wavepacket oscillation model for ultrafast photodynamics

The accurate simulation of photochemical reactions requires methods that capture nonadiabatic transitions through conical intersections between different electronic states. While machine-learning interatomic potentials (MLIPs) offer a promising route to efficient nonadiabatic molecular dynamics, their training for excited states is often limited by the prohibitive cost of generating extensive datasets at a sufficiently high level of quantum chemistry theory. Here, we explore strategies for developing MLIPs that achieve multistate multireference perturbation theory accuracy, systematically comparing single-state, multi-output, multi-state, transfer learning, and delta-learning architectures. We show that transfer learning (TL) from CASSCF to XMCQDPT2 provides the best balance of accuracy and computational efficiency, dramatically reducing the amount of expensive reference data required. This methodology is validated on the methaniminium cation, whose complete photodissociation landscape, including S1 branching into photoisomerization and a direct H2-loss pathway mediated by a recently discovered conical intersection, is captured at the XMCQDPT2/SA(3)-CASSCF(12,12) level. We show that final product yields are largely independent of training strategy, whereas the TL model produces distinct ultrafast population dynamics compared to the randomly initialized model. Finally, we develop a wavepacket oscillation model for fitting excited-state population dynamics, which quantitatively reproduces the ultrafast non-exponential decay and extracts state- and channel-specific lifetimes, directly linking quantum transition probabilities to classical rate constants.