Valence parameters refit
- Pavan Behara
For the simple_fit1, with an interpolated parameter replacing t43, t44, t45, doing a refit of all valence parameters.
Trial 1
Taking current optimized simple_fit1 forcefield as a starting point, which was trained on 170 torsion targets from 14 datasets, and training on the targets in release_1.2.0 tarball with the priors set to 1/10th of their defaults except for the torsions (prior = 2).
Failures with vibfreq targets
Trial 2
Doing two steps of optimization (similar to 1.3.0), where in the first step all the parameters are optimized only with the torsion targets and prior of 2. In the second step on top of these torsion targets the opt-geo and vib-freq targets from 1.2.0 are added.
Here also failures with vibfreq targets
Trial 3
Excluding vibfreq targets from trial 2. Step 1 is a complete refit with only torsion targets and prior 2, step 2 has the opt-geo targets additionally.
Step 1 | Obj. fn (X2) values | interpolated parameters |
|---|---|---|
Only torsion targets prior = 0.2 | 1.65345e+02 → 1.556372e+02 | k1_bondorder1: 9.5933e-01 → 9.3899e-01 k1_bondorder2: 4.1120e+00 → 4.1061e+00 |
Only torsion targets prior = 2 | 1.65345e+02 → 1.462769e+02 | k1_bondorder1: 9.5933e-01 → 9.4165e-01 k1_bondorder2: 4.1120e+00 → 3.9294e+00 |
Taking the prior=2 optimized fit as input for the second step and reducing the prior on all to 1/10th the default, including the torsion parameters. There is negligible change in the parameters and objective function value.
Step 2 | Obj. fn (X2) values | interpolated parameters |
|---|---|---|
Torsion targets + opt_geo targets prior = 0.2 | 2.93328e+03 → 2.933182e+03 | k1_bondorder1: 9.4165e-01 → 9.4158e-01 k1_bondorder2: 3.9294e+00 → 3.9292e+00 |
Since geometry targets are not optimized the result of trial 3 shows a bad performance in RMSD and TFD metrics
SF1_TIG_only: Previous optimized fit with only the newly introduced parameter value optimized
SF1_opts_torsions: Output of trial 3
