Accuracy of advanced neural network potential for reaction thermodynamics of small biomass molecules

Neural network potentials (NNPs) are increasingly used to reproduce quantum-chemical potential energy surfaces at a significantly lower computational cost than conventional electronic-structure methods, yet their accuracy for reaction thermodynamics has not been systematically validated. This study assesses the performance of the PreFerred Potential (PFP) implemented in Matlantis, a recently developed NNP framework, for evaluating reaction thermochemistry in carbohydrate decomposition. Reaction energies and enthalpies are evaluated and compared with density functional theory results and reference values at the level of coupled-cluster with singles, doubles, and perturbative triples at the complete basis set [CCSD(T)/CBS] estimated using a composite method. PFP shows a mean deviation of 1.8 kcal/mol for reaction enthalpies, although notable outliers are observed for reactions involving CO formation.