Test fits

Test fits with linear- and cyclo- alkanes

PhAlkEthOH dataset doesn’t have hessians, so with just the opt geo targets of linear (1 to 7) and cyclo alkanes (3 to 9), separately and both combined optimized H-C-H angle (parameter a2), and the bonds C-C (parameter b1) and C-H (parameter b83) with 90 deg and 1 Ang. as starting values respectively.

Targets:

Here are the final parameter values:

Parameter	Initial values taken	Linear + cyclo alkanes	Cyclo alkanes only	Linear alkanes only	Std. dev. of the three	1.3.0 for reference
b1, length, [#6X4:1]-[#6X4:2]	1.00	1.50	1.50	1.46	0.02	1.52
b1, k, [#6X4:1]-[#6X4:2]	500.00	488.76	487.34	488.01	0.58	517.22
b83, length, [#6X4:1]-[#1:2]	1.00	1.09	1.09	1.09	0.00	1.09
b83, k, [#6X4:1]-[#1:2]	700.00	699.88	699.88	699.86	0.01	754.07
a2, angle, [#1:1]-[#6X4:2]-[#1:3]	90.00	110.51	110.46	110.09	0.19	114.29
a2, k, [#1:1]-[#6X4:2]-[#1:3]	60.00	89.77	90.36	82.16	3.74	66.55

*length in angstroms, angle in degrees, k in kcal/mol/ang^2 or kcal/mol/rad^2

Assigned parameters for each, picked only a2, b1, b83 for fits:

CH4 ['a2', 'b83', 'n16', 'n2']
C2H6 ['a1', 'a2', 'b1', 'b83', 'n16', 'n2', 't3']
C3H8 ['a1', 'a2', 'b1', 'b83', 'n16', 'n2', 't3', 't4']
C4H10 ['a1', 'a2', 'b1', 'b83', 'n16', 'n2', 't2', 't3', 't4']
C5H12 ['a1', 'a2', 'b1', 'b83', 'n16', 'n2', 't2', 't3', 't4']
C6H14 ['a1', 'a2', 'b1', 'b83', 'n16', 'n2', 't2', 't3', 't4']
C7H16 ['a1', 'a2', 'b1', 'b83', 'n16', 'n2', 't2', 't3', 't4']

C3H6 ['a3', 'a4', 'a6', 'b1', 'b83', 'n16', 'n2', 't15', 't16']
C4H8 ['a2', 'a7', 'a9', 'b1', 'b83', 'n16', 'n2', 't2', 't3', 't4']
C5H10 ['a1', 'a2', 'b1', 'b83', 'n16', 'n2', 't2', 't3', 't4']
C6H12 ['a1', 'a2', 'b1', 'b83', 'n16', 'n2', 't2', 't3', 't4']
C7H14 ['a1', 'a2', 'b1', 'b83', 'n16', 'n2', 't2', 't3', 't4']
C8H16 ['a1', 'a2', 'b1', 'b83', 'n16', 'n2', 't2', 't3', 't4']
C9H18 ['a1', 'a2', 'b1', 'b83', 'n16', 'n2', 't2', 't3', 't4']

Another subset with just a1(C-C-C) matching molecules

Removing CH4, C3H6, C4H8 from the set and with rest of the molecules that match to a1

Parameter	Initial values taken	Linear + cyclo alkanes	Cyclo alkanes only	Linear alkanes only	Std. dev. of the three	1.3.0 for reference
*a1, angle, [:1]-[#6X4:2]-[:3]*	100	114.95	113.98	114.70	0.41	113.66
*a1, k, [:1]-[#6X4:2]-[:3]*	90	91.75	96.28	100.85	3.71	99.23

Test fits with sulfonamide (optimizing only a30, a31)

Targets

Torsion drives, opt geo, and vib freq targets that match to a30, a31 from release-1.2.0 targets

Replacing a30 and a31 in openff_unconstrained-1.3.0 with Frostt99 parameters (as starting point) and running the following fits

refit using all the targets (td + opt + vib freq)
refit with only opt_geo targets
refit with only vib_freq targets
refit with only td targets
refit with td+opt_geo targets
test fit with a child parameter for a30
1. a30b *=[#16X4]=* initial value for angle parameter as 125 degrees, and k same as a30 (140)

Fits in progress, most of them crossed five iterations and obj. function values are converging, parameter values at the last iteration as of now (will update after the runs complete)

completed

	Initial values	td+opt+vib	opt	vib freq	td	td+opt	td+opt+vib + child	opt only with child param
	Frostt99	1	2	3	4	5	6	7
a30, angle	109.5	113.0	108.9	115.9	109.9	113.1	113.0	113.1
a30, k	140.0	150.0	171.1	166.9	95.4	139.7	172.2	235.9
a31, angle	109.5	107.5	101.1	114.0	111.4	107.5	110.4	110.0
a31, k	120.0	122.1	153.5	204.2	37.0	117.7	115.9	128.6
a30b, angle	125.0						123.5	125.7
a30b, k	140.0						189.7	180.1
X2 values		1.537827e+03	6.716631e+02	5.480947e+02	2.640341e+02	9.816973e+02	1.391896e+03	5.441547e+02

*angle in degrees, k in kcal/mol/rad^2

Comparison of opt only targets' fits with change in forcebalance options

L2 penalty on the objective function for regularization.

penalty_additive: Factor for additive penalty function in objective function. Add a penalty to the objective function (e.g. L2 or L1 norm) with this prefactor. Using an additive penalty requires an assessment of the order of magnitude of the objective function, but it is closer to the
statistical concept of ridge or LASSO regression.

penalty_multiplicative: Factor for multiplicative penalty function in objective function.
Multiply the objective function by (1+X) where X is this value. Using an multiplicative penalty
works well for objective functions of any size but it is not equivalent to statistical regularization methods.

use_pvals: Bypass the transformation matrix (mvals) and use the physical parameters directly

	Initial values Frostt99	#2 in above table (default option of 1.2.0 fit)
		penalty_additive 1.0	penalty_additive 0.0 penalty_multiplicative -1.0	penalty_additive 0.0 penalty_multiplicative -1.0 use_pvals true
a30, angle	109.5	108.9	112.5	-3.66E-04
a30, k	140.0	171.1	150.6	-4.62E-03
a31, angle	109.5	101.1	105.6	4.26E-04
a31, k	120.0	153.5	123.6	7.75E-05
X2 values		6.716631e+02	6.675653e+02	1.821577e+04