Automating the Development of High-Dimensional Reactive Potential Energy Surfaces with the robosurfer Program System

The construction of high-dimensional global potential energy surfaces (PESs) from ab initio data has been a major challenge for decades. Advances in computer hardware, electronic structure theory, and PES fitting methods have greatly alleviated many challenges in PES construction, but building fitting sets has remained a bottleneck so far. We present the robosurfer program system that completely automates the generation of new geometries, performs ab initio computations, and iteratively improves the PES under development. Unlike previous efforts to automate PES development, robosurfer does not require any uncertainty estimate from the PES fitting method and thus it is compatible with the permutationally invariant polynomial (PIP) method. As a demonstration we have developed five related but different global reactive PIP PESs for the CH3Br + F- system and used them to perform quasiclassical trajectory (QCT) reaction dynamics simulations over a wide range of collision energies. The automatically developed PESs show good to excellent accuracy at known stationary points without any manual sampling, and QCT results indicate the lack of unphysical minima on the fitted surfaces. We also present evidence suggesting that the breakdown of single reference electronic structure theory may contribute significantly to the fitting errors of global reactive PESs.