(WIP) v1.2.0-RC2 fitting result summary

Note: This post will be moved to Force Field / FF releases / Parsley minor release(s) once it is ready to be shared.

Fitting Data and Results

• Fitting targets: 2nd generation training sets (link for the details of the training set generation scheme: http://doi.org/10.5281/zenodo.3777278)

*Note that there are three uncovered torsion parameters(t114, t125, t146) in the torsion training dataset, which are due to the failed torsiondrive calculation carried out inside QCArchive.

• Input force field : version 1.1.0 parsley (http://doi.org/10.5281/zenodo.3695094)

• The objective function decreased from 8.710e+03 to 6.843e+03 in 31 steps.

Benchmark Results

Benchmark data

For the calculation, full benchmark set was used (25168 optimized geometries, plus relative energies of 2005 molecules). Detailed of the molecule selection can be found here: release-1-benchmarking/QM_molecule_selection

(1) Comparison of objective values from single point calculations on benchmark full set

Two types of benchmarks were done to compare the performances: (1) QM vs MM optimized geometries and (2) the relative energies between conformers at “QM optimized geometries”. The final objective function value(X2) from FB single point calculation gives a brief overview of the agreement between QM and MM. The lower X2 is, the better the force field reproduces QM structures and energetics.

(2) v1.2.0-RC1 vs. v1.2.0-RC2: Parameter comparison

• Direct parameter comparison has found no significant differences in equilibrium bond lengths and equilibrium angles while showing notable differences in k values of some angle/ torsion parameters.

• Angle terms with significant different final k values between RC1 and RC2:

  • a6([#1:1]-[*;r3:2]~;!@[*:3], k value in SMIRNOFF99Frosst: 100 kcal/mol/radian²)

  • a3([*;r3:1]1~;@[*;r3:2]~;@[*;r3:3]1, k value in SMIRNOFF99Frosst: 200 kcal/mol/radian²)

  • a15([#8X1:1]~[#6X3:2]~[#8:3], k value in SMIRNOFF99Frosst: 126 kcal/mol/radian²)

: Based solely on my intuition without no strong evidence, RC2 angle k values for the angle terms(doesn’t change much from 1.1.0 during the optimzation) seem physical; 400kcal/mol/radian² for angle seems too large compare to other angle k values. Final gradients for the angle k values are also higher in RC1 (a6: 5.580e+00 , a15 : 4.750e+00) compared to the gradients in RC2 (a6: 1.599e+00 , a15 : 1.391e+00)

*need explanation on the following figures

Here's one example of good examples of having equilibrium angle which is far from angles in QM optimized geometries. Having larger equilibrium angle for a38 helps in locating hydroxyl hydrogen in phosphono group far away enough from its neighboring oxygens, preventing unphysical intermolecular H-bonding.

• Torsion terms with significant different final k values between RC1 and RC2:

  • t146, t147: 6 periodicities

  • t15 ([*:1]-[#6X4;r3:2]-@[#6X4;r3:3]-[*:4]), t16 ([#6X4;r3:1]-[#6X4;r3:2]-[#6X4;r3:3]-[*:4]): in-ring rotation

(3) v1.2.0-RC1 vs. v1.2.0-RC2: Optimized geometries

Specific improvement in certain functional groups(phosphono group, sulfamate acetate) found in RC1 is also shown in RC2.

• TODO: revert WRMSE → objective value

• TODO: implementation of RMSE calculation from FB output

QM optimized geometry of CC(O)([P@@](=O)(O)[O-])[P@](=O)(O)[O-]. ( orange: MM optimized geometry with v1.1.0 force field, green: v1.2.0-RC1 force field, magenta: v1.2.0-RC2 force field)

(4) v1.2.0-RC1 vs. v1.2.0-RC2: Relative energies between conformers at “QM optimized geometries”

Comparison of performances of RC1 and RC2 in reproducing QM relative energies between conformers was carried out. Two different ways to calculate MM relative energies were used. Two different ways to calculate MM relative energies were used. For the left figure, MM relative energies were calculated by taking a difference between MM energy at each point and MM energy at the QM minimum. And for the right figure, MM relative energies were obtained by taking a difference between MM energy at each point and MM energy at QM minimum. Both candidates have smaller MAD and shorter tails than v1.1.0, indicating slight better performances over v1.1.0.