My research aims to make machine learning interatomic potentials a practical tool for studying chemical reactions. I first showed that their differentiability yields accurate analytical Hessians without explicit training, enabling efficient transition-state optimization. Building on this, I address the remaining obstacles to reaction modeling: I developed Popcornn to find reaction paths as continuous neural functions, force field pre-training to keep simulations stable, and large-scale transition-state datasets and benchmarks, including LeMat-TS, to train and evaluate these models.