Search papers, labs, and topics across Lattice.
This paper introduces a fully automated diabatization method using neural networks to fit potential energy matrices (PEMs) for simulating non-adiabatic transitions in molecular photodissociation. The method decomposes the PEM into a zeroth-order diagonal term corrected by a neural network matrix, enforcing symmetry constraints for automatic diabatization. Applied to the photodissociation of CH$_2^+$, the method accurately simulates fragmentation channels and reveals a high cross-section for CH radical formation.
Automating diabatization with neural networks unlocks accurate simulation of complex non-adiabatic molecular dynamics, revealing unexpected fragmentation pathways.
Tracking the complex non-adiabatic transitions in far-ultraviolet photodissociation demands highly accurate diabatic potential energy matrices (PEMs) across numerous excited states. To address this, we introduce a fully automated diabatization method that leverages artificial neural networks to fit PEMs. Our approach divides the PEM into a physically grounded zeroth-order diagonal term, which is then corrected by a neural network matrix to capture electronic couplings. By enforcing symmetry constraints on off-diagonal elements and sharing degenerate diabatic states between the $A'$ and $A''$ irreducible representations, the { diabatization} process becomes completely automatic. We validate this method using time-dependent wavepacket calculations to simulate the photodissociation of CH$_2^+$, incorporating relevant states up to $\approx 13.6$~eV. Finally, we compute partial cross-sections for all fragmentation channels -- including total and partial fragmentation yielding \ce{CH+}, \ce{CH}, \ce{H2}, and \ce{H2+} diatoms -- revealing a notably high cross-section for the formation of the \ce{CH} radical.
