Physics-inspired Inspired Neural Network Models of Reactive Potential Energy Surfaces for Barrierless Reaction Processes

The ab-initio determination of the Potential Energy Surface (PES) for barrierless reactive systems poses significant challenges to theoretical modelling. For such systems, computing tens of thousands of single point energies with multireference methods is standard practice to obtain rate constants applying Variable Reaction Coordinate Transition State Theory (VRCTST)¹, the current state of the art methodology^2,3. Our approach aims at fitting the PES with physics-inspired neural networks (PHINN), in which the PES is mapped as a function of the relative orientation of the reactants. ⁴ The developed methodology has been successfully used to study key reactions for ammonia combustion,⁵ where the PHINN is used as surrogate potential in VRCTST simulations. It was also found that the PHINN PES is sufficiently smooth accurate to perform trajectory simulations. Roaming pathways and reactive channels mediated by long range interactions were in fact determined with accuracy comparable to that obtained using a highly accurate ab initio PES for the H₂CO decomposition process.⁶ It will finally be shown how PHINN trajectories can be used to investigate complicated phenomena, such as the bifurcation of reactive fluxes in the recombination of H with the phenoxy radical.

1.J.Phys.Chem A 107,46,9776-9781,2003

2.Proc.Combust.Inst,vol40,p105270,2024

3.Theor.Chem.Acc.142 118 2023

4.NN-VRCTST.J.Chem.TheoryComput.2025

5.Proc.Combust.Inst.41 105829 2025

6.J.Phys.Chem.A115.50 14370-14381(2011)