Wiggle150: Benchmarking Density Functionals and Neural Network Potentials on Highly Strained Conformers

Accurate benchmarks are key to assessing the accuracy
and robustness
of computational methods, yet most available benchmark sets focus
on equilibrium geometries, limiting their utility for applications
involving nonequilibrium structures such as ab initio molecular dynamics
and automated reaction-path exploration. To address this gap, we introduce
Wiggle150, a benchmark comprising 150 highly strained conformations
of adenosine, benzylpenicillin, and efavirenz. These geometriesgenerated
via metadynamics and scored using DLPNO–CCSD(T)/CBS reference
energiesexhibit substantially larger deviations in bond lengths,
angles, dihedrals, and relative energies than other conformer benchmarks.
We evaluate a diverse array of computational methods, including density-functional
theory, composite quantum chemical methods, semiempirical models,
neural network potentials, and force fields, on predicting relative
energies for this challenging benchmark set. The results highlight
multiple methods along the speed–accuracy Pareto frontier and
identify AIMNet2 as particularly robust among the NNPs surveyed. We
anticipate that Wiggle150 will be used to validate computational protocols
involving nonequilibrium systems and guide the development of new
density functionals and neural network potentials.