A hybrid machine learning–physics model is developed that reduces simulation cost by two orders of magnitude while retaining high ab initio accuracy, to predict free-energy transition states for hydrogen combustion reactions.