1 of 27

Status of theory benchmarking

04/30/21

​

@openforcefield

www.openforcefield.org

2 of 27

2

Jacob’s ladder of chemical accuracy

  • First three semi-local
  • Fourth onwards addition of non-local

Isr. J. Chem. 2020, 60, 787 – 804

www.openforcefield.org

3 of 27

3

QM Benchmark datasets (in general)

  • To address chemically relevant problems:
    • reaction energies
    • barrier heights
    • non-covalent interactions, …
  • Benchmarking against a single set would give the best method for that particular molecular property (biased).
  • Robust method should perform well over a combination of all:
    • Database 2015B: for benchmarking MN15 functional - Donald Truhlar’s group
    • GMTKN55: General Main group Thermochemistry, Kinetics, and Non-covalent interactions - Stefan Grimme’s group
    • MGCDB84: Main Group Chemistry Data Base, 84 subsets - Martin Head-Gordon’s group

​

www.openforcefield.org

4 of 27

4

Database 2015B

  • MN15 benchmark doi.org/10.1039/C6SC00705H

Fig. 1 The percentage of all atomic and molecular databases (AME471), the number after a name means the number of data in this database

www.openforcefield.org

5 of 27

5

GMTKN55 DFT benchmarks

  • General Main group Thermochemistry, Kinetics, and Non-covalent interactions

Lars Goerigk, et al., Phys. Chem. Chem. Phys., 2017, 19, 32184--32215

www.openforcefield.org

6 of 27

6

GMTKN55: Some molecules from Noncovalent interaction datasets (intramolecular)

1. SCONF: Relative energies of sugar conformers

2. BUT14DIOL: Relative energies in �butane-1,4-diol conformers

3. UPU23: Relative energies between RNA-backbone conformers

www.openforcefield.org

7 of 27

7

GMTKN55: Some molecules from Noncovalent interaction datasets (intramolecular)

4. IDISP: Intramolecular dispersion interactions

5. MCONF:

Relative energies in melatonin conformers

6. ICONF: Relative energies in conformers of inorganic systems

www.openforcefield.org

8 of 27

8

GMTKN55: Some molecules from Noncovalent interaction datasets (intramolecular)

7. PCONF: Relative energies in tri- and tetrapeptide conformers

8. AMINO20x4: Relative energies in amino acid conformers

and others

www.openforcefield.org

9 of 27

9

MGCDB84

  • Main Group Chemistry Data Base - doi.org/10.1080/00268976.2017.1333644

www.openforcefield.org

10 of 27

10

MGCDB84: Some molecules from isomerization datasets

www.openforcefield.org

11 of 27

11

MGCDB84: Some molecules (monomers) from Noncovalent interaction datasets

3B-69-TRIM: Binding energies of trimers, with three different orientations of 23 distinct molecular crystals

Uracil

www.openforcefield.org

12 of 27

12

XB51: Binding energies of large halogen-bonded dimers

MGCDB84: Some molecules from Noncovalent interaction datasets

www.openforcefield.org

13 of 27

13

MGCDB84: Some molecules from Noncovalent interaction datasets

CO2Nitrogen16: Binding energies of CO2 to molecular models of pyridinic N-doped graphene

​

www.openforcefield.org

14 of 27

14

MGCDB84: Some molecules from Noncovalent interaction datasets

Ionic43: Binding energies of anion-neutral, cation-neutral, and anion-cation dimers

www.openforcefield.org

15 of 27

15

MGCDB84: Some molecules (monomers) from Noncovalent interaction datasets

AlkBind12: Binding energies of saturated and unsaturated hydrocarbon dimers

www.openforcefield.org

16 of 27

16

Other relevant studies

  • Some other recent benchmarks

JCTC. 2018, 14, 1254−1266

J. Chem. Inf. Model.2017, 57, 1265−1275

J. Chem. Theory Comput. 2016, 12, 444−454

Isr. J. Chem. 2020, 60, 787 – 804

J. Phys. Chem. Lett. 2021, 12, 11, 2796–2804

J. Chem. Theory Comput. 2021, XXXX

www.openforcefield.org

17 of 27

17

Functionals to try

  • PW6B95-D3(0)/D3BJ:
    • not considered in MPCONF196
    • CHAL336 - among top3 hybrids for overall set
    • comparable computation time with B3LYP-D3BJ
    • slight improvement over B3LYP-D3BJ in CHAL336 (0.95 vs 0.72 kcal/mol mean abs. dev.)
    • good performance on GMTKN* benchmark sets
    • not recommended in MGCD84
  • M06-2X-D3(0):
    • MPCONF196 study is against it but CHAL336 study shows better than B3LYP performance
  • M05-2X-D3(0): Kieron Burke’s recent torsion paper shows a better performance
    • closer to density-corrected-DFT
  • M08-HXD3(0): In general, M-* functionals show good performance on GMTKN55 set

​

www.openforcefield.org

18 of 27

18

Functionals to try

  • ⍵B97M-V: no analytical gradients for VV10 (pending since 2018), highly expensive as of now
    • ⍵B97M-D3BJ version is available, second only to VV10 version in GMTKN55 set hybrids
    • worse in chalcogen−halogen interactions but best hybrid in CHAL336 study
    • no need of density corrections as in b3lyp
    • introduced in 2016 and was missing in most studies beyond GMTKN55/MGCDB84
  • ⍵B97X-D3: Recommended in both GMTKN55 and MGCDB84 sets
    • ANI* models are trained on data that use ⍵B97X/6-31G(d)
    • overall performance is not better than B3LYP-D3BJ in CHAL336 study
    • MPCONF196 study skips this due to convergence issues (may not be functional related/technical)
  • Other versions of B3LYP:
    • B3LYP-NL
    • B3LYP-D3MBJ (refit version of D3)

​

www.openforcefield.org

19 of 27

19

Double hybrids - Strict NO!

  • Basis set dependant and typically needs quadruple-zeta level basis - expensive!!!
  • Some double hybrids are not present in libxc (Susi Lehtola recently merged, we may see them in psi4-1.4 soon)
  • No analytical gradients for double hybrids and range-separated hybrids with VV10 in psi4 yet
    • MP2 analytical derivatives are already available but double hybrids aren’t updated yet, on psi4 list of things but not sure of timeline
    • same goes for range separated hybrids, hold up related to one last step
  • Orca - an alternative? (NO! from SB+JW)

www.openforcefield.org

20 of 27

20

Baseline to compare

  • level1/basis1 // level2/basis2
    • optimize geometry at lower level and run a single point energy evaluation at higher level
    • CCSD(T)/CBS // MP2/6-311+G*
    • DLPNO-CCSD(T)/CBS // HF/6-31G*
    • MP2/CBS // HF/6-31G*
  • Need to find which is a better option

www.openforcefield.org

21 of 27

21

Conclusion

  • Need to find the sweet spot between computation time and accuracy
    • benchmarks are mostly done at larger basis (def2-QZVPP basis in GMTKN55)
    • transferability of various functionals to smaller basis sets
    • conformational energies and torsional profiles
  • Have to submit few more compute specs.

​

​

​

​

​

  • Thinking of picking up three basis sets and submit for all functionals, any feedback?
    • 58 torsiondrives submitted by Hyesu
    • can use MPCONF196 for conformational energies

Method

B3LYP-D3BJ

B3LYP-D3BJ

B3LYP-D3BJ

B3LYP-D3BJ

B3LYP-D3BJ

B3LYP-D3BJ

B3LYP-D3BJ

B3LYP-D3BJ

Basis

DZVP

def2-

tzvp

def2-

tzvpd

def2-

tzvpp

def2-

tzvppd

def2-

qzvp

6-31+

g**

6-311+

g**

#targets in 1.2.0

​

Geo opt

240

Vib freq.

1189

Torsion profiles

710

www.openforcefield.org

22 of 27

22

  • TG’s Q: Are there edge cases or bad behavior with current level of theory that prompted this work?

Ans: Motivation for theory benchmarking work (in Hyesu’s words)

    • We do not know how well DZVP performs for charged molecules, and for properties other than conformational energies;
    • DZVP results are less accurate than literature for medium/large systems largely due to lack of DZVP-specific dispersion parameters and three-body term. (although the difference is small)

Discussion

www.openforcefield.org

23 of 27

23

  • Only for uncertainty in conformer energies
  • Doesn’t have more modern functionals (wB97, M0x variants)
  • Doesn’t answer which level is good
    • for charged molecules (dipole moments)?
    • for torsional barriers?

Victoria’s work

www.openforcefield.org

24 of 27

24

Interesting chemistries

  • Detour: Any interest in adding some Selenium compounds?

www.openforcefield.org

25 of 27

Status update

07/02/2021

25

@openforcefield

www.openforcefield.org

26 of 27

26

QM Data generation

  • Added three more compute specs to the analysis
    • wB97X-D3BJ, PW6B95-D3(0), B3LYP-D3MBJ
  • Constrained optimizations
    • updated MP2/heavy-aug-cc-pvtz to include the augmented set
    • wB97X-V, B3LYP-NL - both don’t have analytic gradients for dispersion, and more failures (thinking of taking them out of the study)
  • Few geometric convergence failures in some of the torsiondrives with larger basis
    • TZVPD, TZVPP - debugging locally
      • keyword changes didn’t work: ‘tric’ to ‘dlc’, increased maxiter
      • need to look deeper
  • Improving documentation [WIP]:

www.openforcefield.org

27 of 27

27

Analysis scripts [WIP]

  • Most time consuming parts in coding are preprocessing and the right visualization
  • ​

www.openforcefield.org