QM Theory Benchmark

Driver	Approver	Contributors	Stakeholder
Pavan Behara Hyesu Jang

Objective	Benchmark QM theory for FF training data
Due date	30 Sep 2021
Key outcomes	To answer some questions: How well DZVP performs for charged molecules, and for properties other than conformational energies, how large our basis should be, how many diffuse functions? How well does current level of theory estimate torsion barriers, especially when noncovalent interactions dominate? Compare with other levels of theory being used in building other force fields like ANI.
Status	IN PROGRESS

Problem Statement

A wide range of density functional methods and a large number of basis sets are available to derive the electronic structure and properties of molecules. Estimating the accuracy of calculation method for desired properties, and choosing a method with a low computational cost, need a comprehensive evaluation of the methods on a test set of representative molecules. Conformational energies and a set of torsion profiles have been benchmarked before by Lee-Ping (link) and B3LYP-D3BJ/DZVP level of theory has been chosen for building OpenFF force fields. There are still some unanswered questions on

whether this is a good choice for charged molecules?
any density-driven or delocalization errors creep up in torsion scans as shown in Burke’s papers (https://dx.doi.org/10.1021/acs.jpclett.1c00426 )
how good it is in describing problematic chemistries, like
- biaryl torsion barriers (http://dx.doi.org/10.1021/ct5004725 ).
- chalcogen, halogen interactions (http://doi.org/10.1021/acs.jctc.1c00006 ). Strong polarizability from these elements induce strong dispersion interactions, and there is a need to check on how the current level of theory is performing and whether the dispersion corrections are good enough to compensate for that.
- and non covalent interactions in general.
transferability of accuracy shown in density functional benchmarks with larger basis sets (from literature) to smaller basis sets?

Scope

Must have:

Accuracy of torsional profiles, accuracy of dipole and quadrupole moments, comparison of computational performance of methods

Nice to have:

Specific chemistry insights viz., molecules containing iodine, bromine, selenium, arsenic (roxarsone, tetraphenylarsonium, …)

Not in scope:

Timeline

Milestones and deadlines

Milestone	Owner	Deadline	Status
Initial set of molecules (selection criteria for molecules, selection criteria for additional mols, qca-dataset-location)	Hyesu Jang		COMPLETED
Dataset tracking initial set of torsion scans with B3LYP-D3BJ and a variety of basis sets (PR link) - few failed calculations, checking with local runs constrained minimizations with the final optimized geometries at higher level theory and tzvpd basis (PR link) - need to augment from 36 to 59 B3LYP-NL and wB97X-V both don’t have analytic gradients for dispersion terms, so we may ignore these specs?? additional compute specs with other DFT functionals (wB97X-D3BJ, PW6B95-D3) with similar or better accuracy (PR link) - running now	Pavan Behara		IN PROGRESS
Analysis scripts for torsions comparison, dipole, quadrupole comparisons (repo link)	Pavan Behara	09 Jul 2021	IN PROGRESS
Submit an initial written review and propose any additional work to be done	Pavan Behara	13 Aug 2021	NOT STARTED

Reference materials

Lee-Ping’s benchmark of conformational energies on MPCONF196 set and a select set of 15 torsion profiles (link), and discussion about DZVP (link)
Lee-Ping’s benchmark with the new basis set configuration that addresses the Iodine issue raised by Bill Swope (link).
Some initial summary on current DFT benchmarks from literature (link)