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

Area of Science:

• Computational Chemistry
• Quantum Chemistry
• Method Development

Background:

• Existing computational benchmarks primarily focus on equilibrium geometries.
• This limits the assessment of methods for nonequilibrium structures crucial in molecular dynamics and reaction path exploration.
• There is a need for benchmarks representing highly strained molecular conformations.

Purpose of the Study:

• Introduce Wiggle150, a novel benchmark set of 150 highly strained molecular conformations.
• Evaluate the performance of diverse computational methods in predicting relative energies for these challenging geometries.
• Identify methods suitable for applications involving nonequilibrium molecular structures.

Main Methods:

• Generated 150 highly strained conformations for adenosine, benzylpenicillin, and efavirenz using metadynamics.
• Calculated reference energies using high-level composite quantum chemical methods (DLPNO-CCSD(T)/CBS).
• Assessed various computational approaches: density-functional theory (DFT), composite methods, semiempirical models, neural network potentials (NNPs), and force fields.

Main Results:

• Wiggle150 geometries exhibit significantly larger deviations in structural parameters and relative energies compared to existing benchmarks.
• Multiple computational methods demonstrate a favorable speed-accuracy trade-off on the Wiggle150 set.
• AIMNet2 emerged as a particularly robust neural network potential for predicting relative energies of strained conformations.

Conclusions:

• Wiggle150 serves as a valuable tool for validating computational protocols for nonequilibrium systems.
• The benchmark can guide the development of improved density functionals and neural network potentials.
• This work advances the accuracy and robustness assessment of computational chemistry methods.