Theoretical Uncertainties

4) Theoretical Uncertainties

This section controls perturbation model, uncertainty magnitudes, distributions, and convergence thresholds for parameter classes.

4.1 Global perturbation and score-balance keys

Keyword	Default value	Description
max_std	4	Maximum deviation from the initial model, in standard deviations. For multiplicative parameters this is a log-space bound: the trusted range is value * uncertainty**(±max_std) (i.e. [value / uncertainty**max_std, value * uncertainty**max_std]), so the theory score exactly reaches max_std**2 at the boundary.
freq_mode	"batch"	Frequency perturbation mode. Possible values: "batch" or "individual". Individual mode has not been thoroughly tested.
weight_theory	1.0	Raw theory contribution weight; normalized at runtime with weight_experiments.
weight_experiments	1.0	Raw experiment contribution weight; normalized at runtime with weight_theory.
specific_std	{}	Per-parameter override map for standard deviations. An entry overrides the global std_<ptype> for that parameter and governs the sampling scale, the scoring weight, and the perturbation boundaries (trusted range): a larger value widens the accepted bounds, a smaller one tightens them. Parameters without an entry fall back to the global std_<ptype>.

4.2 Standard deviations (std_*)

Keyword	Default value	Description
std_we	1.0	Well and bimolecular fragments energy uncertainty (kcal/mol).
std_be	1.5	Barrier energy uncertainty (kcal/mol).
std_ifc	1.1	Individual vibrational frequency multiplicative uncertainty. Multiplicative factor.
std_bfc	1.05	Batch vibrational frequency multiplicative uncertainty. Multiplicative factor.
std_hrs	0.1	Hindered rotor uncertainty. Percentage.
std_if	1.1	Imaginary frequency multiplicative uncertainty. Multiplicative factor.
std_fact	0.25	Energy transfer factor uncertainty. Percentage.
std_pow	0.075	Energy transfer power uncertainty. Percentage.
std_epsilon	0.1	Lennard-Jones epsilon uncertainty. Percentage.
std_sig	0.1	Lennard-Jones sigma uncertainty. Percentage.
std_sfc	2.0	Symmetry-factor uncertainty for barrierless reactions. While the symmetry has no uncertainty, this parameter allows to scale the state density. Multiplicative factor.
std_mrc	1.5	Multi-dimensional rotor symmetry uncertainty. While the symmetry has no uncertainty, this parameter allows to scale the state density. Multiplicative factor.

Distribution restrictions. Each parameter belongs to a category that constrains which sampling distribution (distrib_*) it may use:

• Multiplicative parameters (if, sfc, mrc, bfc, and the individual/batch frequencies) accept only log-normal or log-uniform.
• Additive (we, be, pow) and percentage (hrs, sigma, epsilon, fact) parameters accept only uniform or normal.

Any other combination cancels the run. In the GUI, only the valid distributions for a parameter's category are selectable.

For multiplicative parameters the theory-score distance is measured in log space as ln(value / reference) / ln(uncertainty), so the penalty is symmetric under a factor and its inverse and consistent with the log-normal (log-space) perturbation of these parameters. Additive and percentage parameters keep their linear distance.

The perturbation, trust boundaries, and derivative steps for multiplicative parameters are all computed in log space consistently with this scoring:

• The trusted range is value * uncertainty**(±max_std), so the theory score at the boundary equals max_std**2 exactly.
• The log-normal sampling uses a log-space sigma of ln(uncertainty), giving ±2/3/4 sigma coverage of 95.45% / 99.73% / 99.99% of samples within the corresponding value * uncertainty**(±k) range.

Note: with this log-space reconciliation the multiplicative trust region is wider than in earlier versions for uncertainty > 1; multiplicative-run results are not bit-for-bit comparable with pre-reconciliation runs. Additive and percentage parameters are unaffected.

4.3 Convergence thresholds (conv_*)

Keyword	Default value	Description
max_score	4.0	Convergence threshold. Maximum score for the best-model ensemble has to be lower.
score_conv	2	Convergence threshold. Average score for best-model ensemble has to be lower.
param_conv	0.01	Parameter-space convergence threshold used for both parameter values and standard deviations. For additive and percentage parameters it is a percentage (relative change), computed from the arithmetic per-generation mean/std. For multiplicative parameters (if, sfc, mrc, bfc, frequencies) the per-generation statistics are the geometric mean and geometric standard deviation, and genetic-algorithm convergence is measured in log space as `
conv_we	0.1	Convergence threshold for well energies.(kcal/mol)
conv_be	0.1	Convergence threshold for barrier energies.(kcal/mol)
conv_pow	0.01	Convergence threshold for energy transfer power.