Abstract
<jats:p>Building reaction networks for kinetic modeling hinges on locating transition states (TSs) in order to compute reaction rates. In many cases, the reactions of interest involve complex, multi-step mechanisms with intermediates. Double-ended methods such as the nudged elastic band (NEB) are built for elementary steps, where a single TS separates reactant and product, and they struggle (or even fail) when given endpoints that straddle several saddle points, corresponding to multi-step mechanisms. We present an autosplitting NEB (ASNEB) algorithm that detects nascent intermediates during path optimization, splits the reaction at those minima, and recurses on the resulting segments until every step is elementary. Network completions then probe alternative connections among the discovered reactant, intermediates, and product, exposing competing mechanisms. Running discovery at a semiempirical level (GFN2-xTB) and refining intermediates and TSs with DFT keeps the cost dominated by the same ab initio refinement already required by conventional modeling, while improving TS guess quality and lowering the computational cost. We demonstrate the utility of ASNEB with the Oxy-Cope, Wittig, and Dakin reactions, where ASNEB recovers the established mechanisms and uncovers competing pathways with no prior specification of intermediates.</jats:p>