Duo Fitting¶

Duo allows the user to modify (refine) the potential energy curves and other coupling curves by least-squares-fit to experimental energy term values or wavenumbers.

Duo uses Gauss-Newton least-squares fitting supplemented with the Marquardt approach. Optionally, this algorythm can be supplemented by a linear search via the keyword linear-search. The associated system of linear equations is solved either using the LAPACK subroutine DGELSS or a home-made subroutine LINUR as controlled by the keyword FIT_TYPE.

Fitting is, by far, the trickiest part of Duo, both on the part of the program itself and on the part of the user. While the calculation of energy levels and spectra from given PECs, couplings and dipole curves is relatively straighforward (the most critical point being the consistency of the phases specified for the various coupling terms), fitting is often more difficult and may require a trial-and-error approach. Fitting is also the part of Duo where most improvements are to be expected in future new versions.

Example of a fitting section:

FITTING
JLIST 2.5,0.5, 1.5 - 11.5, 22.5 - 112.5
itmax 30
fit_factor  1e6
output alo_01
fit_type dgelss
fit_scale 0.001
linear_search 5
lock      5.0
robust      0.001
energies   (J parity NN  energy ) (e-state v ilambda isigma omega  weight)
 0.5   +   1     0.0000  1  0  0   0.5   0.5    100.000
 0.5   +   2   965.4519  1  1  0   0.5   0.5      7.071
 0.5   +   3  1916.8596  1  2  0   0.5   0.5      5.774
 0.5   +   4  2854.2366  1  3  0   0.5   0.5      5.000
 0.5   +   5  3777.5016  1  4  0   0.5   0.5      4.472
 0.5   +   6  4686.7136  1  5  0   0.5   0.5      4.082
 0.5   +   7  5346.1146  2  0  1  -0.5   0.5    100.000
end

When state labels are used as potential IDs, i.e.

potential X
  ......
  ...
end

potential A
  ......
  ...
end

The same labels must be used to identify the electronic states in the energy part of the FITTING section:

FITTING
JLIST 2.5,0.5, 1.5 - 11.5, 22.5 - 112.5
itmax 30
fit_factor  1e6
lock      5.0
energies   (J parity NN  energy ) (e-state v ilambda isigma omega  weight)
 0.5   +   1     0.0000  X  0  0   0.5   0.5    100.000
 0.5   +   2   965.4519  X  1  0   0.5   0.5      7.071
 0.5   +   3  1916.8596  X  2  0   0.5   0.5      5.774
 0.5   +   6  4686.7136  X  5  0   0.5   0.5      4.082
 0.5   +   7  5346.1146  A  0  1  -0.5   0.5    100.000
end

Fitting ranges¶

The fitting ranges can be proved to set bound for the fittnig parameters as follows:

poten Ap
name "Ap2Delta"
lambda 2
mult   2
type   EHH
values
V0           1.47076002476226e+04  fit
RE           1.81413234392416e+00  fit  range 1.7,1.9
DE           5.92200000000000E+04       link 1,2,3
ALPHA        1.51288898501269E+00  fit
C            6.54666674267795E-03  fit
B1          -2.63874034064552E+00  fit
B2          -4.80709545680641E+00  fit
B3           0.000000000000000000
end

where the keyword range is followed by the corresponding bound for the corresponding parameter (using the same units). The keyword range and fit are complementary. The input is not sensitive to their order, as long they appear after column 2.

These bound will prevent the varying parameter $f_{i-}\to f_{i}$ at the iteration $i$ to get a value outside its bounds by applying the following conditions: .. math:

\begin{split}
 f_{i} &= \max(f_{i-1},f_{\rm min} ) \\
 f_{i} &= \min(f_{i-1},f_{\rm max} ) \\
\end{split}

where $f_{\rm min}$ and $f_{\rm max}$ are the corresponding bounds.

Keywords¶

FITTING

This keyword marks the beginning of the fitting input section. The whole section can be deactivated by putting none or off next to the keyword FITTING. This is useful to disable the fitting without removing the input block from the input file.

jlist (aliases are jrot and J)

This keyword allows the user to specify the values of the $J$ quantum number to be used in the fit. It superseedes the corresponding jrot keyword specified in the general setup Individual values of $J$ can be separated by spaces or commas, while ranges are specified by two values separated by a hyphen (hyphens should be surrounded by spaces). The first $J$ value is used to determine ZPE. For example

JLIST  1.5, 5.5, 15.5 - 25.5, 112.5

selects the values 1.5, 5.5, all values from 15.5 to 25.5 and the value 112.5.

itmax (alias itermax) An integer defining the maximum number of fitting iterations.

Setting itmax to zero implies that no fit will be performed (straight-through calculation); however, the differences between the computed energy levels (or frequencies) and the reference (experimental) ones will be printed. Example:

itmax 15

fit_factor This factor is used when reference curves of the
abinitio type are included in the fit and used to define the importance of the energy/frequency data relative to the reference abinitio data. This factor is applied to all energy (frequencies) weight factors $w_i^{\rm en}$ .

When the factor is very large (e.g. $10^6$ , like in the example above) the penalty for deviating from the reference ‘ab initio’ curve is very small, so that only the obs. - calc. for energy levels matter. Vice versa, if the factor is very small (e.g. $10^{-6}$ ) the fit is constrained so that the fitted curves stay very close to the reference (abinitio) ones. When this number is extremely small (smaller than $10^{-16}$ ) the experimental data are completely ignored and the fit is performed to the aivalues only. Thus this feature also allows one to use the FITTING section for building analytical representations (see type-s currently available) of different objects by fitting to the corresponding aior reference data provided in the abinitio-sections of the input.

Example:

fit_factor 1e2

Lock,``Thresh_Assign``

Lock or Thresh_Assign denotes the threshold (cm^-1) for which the quantum numbers are used to match to the experimental value and to lock during the refinement. The quantum numbers defining state, $v$ , $|\lambda|$ , $|\sigma|$ and $|\Omega|$ will be used to identify, match and lock the energy value in place of the running number within the $J$ /parity block. When Lock is zero or not present, this feature is switched off and the running number is used to match the experimental and calculated values. When negative, the match is reconstructed based solely on the closest value within the lock-threshold given. If the match within the lock-region is not found, the row $J$ /parity number is used to match the theoretical and experimental energies. For example to match and lock to the calculated energy to the experimental one based on the quantum numbers within 20 cm^-1 use:

lock 20.0

thresh_obs-calc This keywords triggers switching off states from the fit if the obs.-calc. residuals become larger than the threshold specified.

This feature is useful in case of multiple swapping of the states during the fits and even the lock option does not help. The default value is zero (the feature is off).

range This keyword is to specify the fitting bounds for a given fitting parameter and should appear on the same line as the parameter value, normally after the keyword fit:

poten Ap
.....
....
values
V0           1.47076002476226e+04  fit
RE           1.81413234392416e+00  fit  range 1.7,1.9
......
end

robust This keyword allows the user to switch on

Watson’s robust fitting procedure: 0 is off, any other positive value is on and defines the target accuracy of the fit as given by the weighted root-mean-square error. The robust-value is the targeted accuracy (obs.-calc.) of the fit. Example:

robust 0.01

target_rms

This is to define the convergence threshold (cm^-1) for the total, not-weighted root-mean-squares (rms) fitting error. Example:

target_rms 0.1

output

This is the filename for the files name.en, name.freq and name.pot, containing detailed information on the fitting, including the fitting residuals for each iteration. Example:

output NaH_fit

linear_search

When the linear_search (Damped Gauss-Newton) keyword is given and the associated value integer value is not zero, Duo will attempt a linear search of the scaling factor $\alpha$ for the correction parameter vector ${\bf x}$ :

${\bf x}_{i+1} = {\bf x}_{i} + \alpha \Delta {\bf x}$ ,

where ${\bf x}_{i}$ is the paramor vector for the iteration $i$ , $0 \le \alpha \le 1$ , $\Delta {\bf x}$ is the least-squares correction. The step length $\alpha$ needs to satisfy the Armijo condition. The keywords linear_search comes with an integer parameter $N$ defining the maximal number of steps in the linear search staring from $\alpha = 1, 1 - 1/N, 1 - 2/N \ldots$ . Example:

linear_search 5

fit_type

There are two linear solvers available to solve the linear systems associated with the noon-linear squares fit using Gauss-Newton, fit_type LINUR (home made solver) and fit_type DGELSS (LAPACK). The latter should be more stable for strongly correlated systems while LINUR is usually with faster convergence, but for most cases these two methods should be equivalent.

fit_scale

This is fixed-value analogy of the linear scaling. It directly defies a scaling factor $\alpha$ used to scale the parameter vectors increment :math:Delta {bf x}`, see above. It is ignored when linear_scaling is defined. It can be used to improve the convergence.

Example:

fit_scale 0.5

energies

This keyword starts the section with the energy levels to be fit to (e.g., obtained from an analysis of the experimental line positions). Energy levels are written as in the following example:

energies
   0.5   +    1     0.0000 1  0 0 0.5  0.5  1.00
   0.5   +    2   965.4519 1  1 0 0.5  0.5  0.90
   0.5   +    3  1916.8596 1  2 0 0.5  0.5  0.80
end

where the meaning of the various quantities is as follows; col.1 is the total angular momentum quantum number $J$ ;

col. 2 either the total parity $\tau = \pm$ or the $e/f$ parity;

col. 3 is a running number $N$ couting levels in ascending order of the energy within a $(J,\tau)$ symmetry block;

col. 4 is the energy term value $\tilde{E}$ , in cm^-1;

col. 5 is the electronic state index state, as labelled in the potential sections, e.g. 1 or X;

col. 6 is the vibrational quantum number $v$ ;

col. 7 is the projection of the electronic angular momentum $\Lambda$ for the state in question (an integer); only the absolute value is used to match to the calculated value;

col. 8 is the projection of the total electronic spin $\Sigma$ (integer of half integer); only the absolute value is used to match to the calculated value;

col. 9 is the projection of the total angular momentum $\Omega$ (integer of half integer); only the absolute value is used to match to the calculated value;

col. 10 is the weight $W$ of the experimental energy in question (a real and positive number usually given by $\sigma^{-2}$ , where $\sigma$ is the uncertainty of the energy level).

frequency (aliases are frequencies and wavenumbers)

This keyword works similarly to the energies keyword above but starts the section specifying the wavenumbers (i.e., line positions) to be fitted to.

Example:

frequencies
  0.0  +   2 0.0 +  1   720.0000   2  0   1  -1.0   0.5    1  0   0   0.0   0.0  1.00
  2.0  +  17 3.0 -  2  5638.1376   4  0   0   1.0   1.0    2  0  -1  -1.0  -2.0  1.00
  4.0  +  17 5.0 -  2  5627.5270   4  0   0   1.0   1.0    2  0  -1  -1.0  -2.0  1.00
  4.0  +  18 7.0 -  2  5616.7976   4  0   0   0.0   0.0    2  0  -1  -1.0  -2.0  1.00
end

The meaning of the quantities in each line are the following (see the keyword energies above for an explanation of the symbols. The prime/double prime symbol correspond to upper/lower level): $J'$ , $\tau'$ , $N'$ , $J''$ , $\tau''$ , $N''$ ; frequency (cm^-1); state $'$ , $v'$ , $\Lambda'$ , $\Sigma'$ , $\Omega'$ ; state $''$ , $v''$ , $\Lambda''$ , $\Sigma''$ , $\Omega''$ ; weight.

off, none are used to switch off Fitting, Intensity or Overlap, when put next to these keywords.

Structure of the fitting output¶

During fitting Duo will print for each iterations the fitting residuals using the following structure (the first line with numbers 1 to 20 is not part of the output but serves as a legend):

2     3   4           5           6          7         8   9    10   11    12    13    14  15  16    17    18    19    20

1    0.5  +      0.0000      0.0000     0.0000  0.60E-02   1     0    1  -0.5   0.5   0.5   1   0     1  -0.5   0.5   0.5
2    0.5  +   1970.2743   1970.3983    -0.1240  0.59E-02   1     1    1  -0.5   0.5   0.5   1   1     1  -0.5   0.5   0.5
3    0.5  +   3869.6639   3869.7934    -0.1295  0.30E-02   1     2    1  -0.5   0.5   0.5   1   2     1  -0.5   0.5   0.5
4    0.5  +   5698.7392   5699.2951    -0.5559  0.20E-02   1     3    1  -0.5   0.5   0.5   1   3     1  -0.5   0.5   0.5
1    0.5  -      0.1001      0.0000     0.1001  0.60E-02   1     0   -1   0.5  -0.5   0.5   1   0    -1   0.5  -0.5   0.5
2    0.5  -   1970.4156   1970.3983     0.0173  0.59E-02   1     1   -1   0.5  -0.5   0.5   1   1    -1   0.5  -0.5   0.5

The meaning of the quantities in the various columns is as follows;

col.1 is a simple line counter $i$ counting over all lines;

col.2 is a counter $N$ counting lines within each $J, \tau$ symmetry block;

col. 3 is $J$ ; col. 4 is the parity $\tau$ ;

col.5,6 are, respetively, the reference (Observed) and the calculated value of the line position;

col.7 is the difference between observed and computed line positions;

col. 8 is the weight assigned to the transition in the fit;

col. 9 to 14 are the quantum numbers of the lower state: state, $v$ , $\Lambda$ , $\Sigma$ , $\Omega$ and $S$ ;

col. 15 to 20 are the quantum numbers for the upper state (same definition as for columns 9 to 14).

The auxiliary files .en, .freq, .pot

The files name.en contains all computed term values together with the theoretical quantum numbers, compared to the experimental values, when available, along with the experimental quantum numbers as specified in the fitting section, for all iterations of the least-squares fit. Here name is the file name as speficied by the output keyword. The output is in the same format as in the standard output file (see above) with the difference that it contains all calculated values (subject of the nroots keyword, see Section ref{s:diagonaliser}). An asterisk * at the end of the line indicates that either the theoretical and experimental assignments don’t agree or a residuals obs.-calc. is too large (large than the lock parameter).

The frequency file name.freq with the keyword frequencies. It has a similar structure as the standard output, with the difference that for each transition from the frequency section the program will estimate additional transition frequencies involving energies (both lower and upper) which are within lock cm^-1 of the corresponding input values. This is done to facilitate the search for possible miss-assignment, which is typical for transitions. This is printed out for all iterations.

The file {name.pot (potential) contains the residuals between the fitted and the reference curve (if specified by an abinitio object). The file is overwritten at each iteration.

Fitting grid points¶

Although usually the fitted object has to be represented in some parameterised functional with the parameters varied in a least-squares fit. It is however also possible to vary grid values of a field in the grid representation. The idea behind this approach is based on the fact that Duo uses cubic splines when mapping a (usually smaller) grid ( $N_{\rm field}$ ) of values given in the input to the (usually larger) Duo grid of points $N_{\rm Duo}$ . In principle, the functional values $f_i(r_i)$ can be treated as parameters defining the function $f(r)$ via its values at the corresponding grid points $r_i$ ( $i=1..N_{\rm field}$ ) via the cubic splines. This works especially well for a very small number of points $N_{\rm field}$ . In Duo input is very analogous to the standard parameterised fit via the keyword fit after the parameter values, for example

spin-orbit  1 1
name   "<X2Delta|SO|X2Delta>"
spin   0.5 0.5
lambda  2 2
sigma  0.5 0.5
type  grid
factor  1.0  (sqrt2)
values
0.750000000000       -597.1915000     fit
1.200000000000       -590.2260000     fit
1.500000000000       -599.7850000     fit
2.000000000000       -603.9980000     fit
3.000000000000       -603.0000000     fit
end

Constrained fit to ab initio¶

If abinitio fields are provided, the is automatically constrained to the ab initio values of a given field. This can be useful, when the number of experimental data is very limited in order to be able all the parameters required to represent the full complexity of the fitted function. This is implemented as a simultaneous of the same parameters both to the experimental energies (frequencies) and to the ab initio values. In this case it is necessary to control the relative importance of the experimental and the ab initio data using the keyword fit_factor in the fitting section, for example:

::

FITTING JLIST 2.5,0.5, 1.5 - 11.5, 22.5 - 112.5 itmax 30 fit_factor 1e6 …..

The (real) value of fit_factor $s$ is a factor used to scale the experimental fitting weights $w_i^{\rm exp}$ . The ab initio weight factors can be also scale individually also using the keyword fit_factor placed in the corresponding ``abinitio field, for example:

abinitio poten X
name "X2Sigma"
lambda 0
symmetry +
mult   2
fit_factor 0.001
type   grid
values
1.400000       40782.9118
1.450000       28672.6462
....
end

As described in the section about the fitting keywords, when the ‘experimental’ fit_factor is very large (e.g. $10^6$ ) the penalty for deviating from the `abinitio` (reference) curve is very small, so that only the `obs. - calc.` for energy levels matter. Vice versa, if the experimental factor is very small (e.g. :math:`10^{-6}`) or if the '*ab initio*' ``fit_factor is very larger, the fit favours the reference (abinitio) data. When this experimental fit_factor is extremely small (smaller than $10^{-16}$ ) the experimental data are completely ignored and the fit is performed to the ab initio values only. Thus this feature also allows one to use the FITTING section for building analytical representations (see type-s currently available) of different objects by fitting to the corresponding aior reference data provided in the abinitio-sections of the input.

It should be noted that the reference (abinitio) curve does not have to be a grid field. Any representation, including analytic ones cane be used to constrained the varying parameters to a reference field.

Ab initio weights¶

For the constrained fit to work, the ab initio data must be weighted. This is done by adding a column 3 with the corresponding weights, e.g. :

abinitio poten X
name "X2Sigma"
lambda 0
symmetry +
mult   2
type   grid
fit_factor 0.1
values
1.400000       40782.9118   0.5
1.450000       28672.6462   0.6
1.500000       19310.0950   0.7
1.550000       12244.8264   0.8
1.600000        7101.7478   0.9
1.650000        3562.3190   1.0

Alternatively, the weights can be also defined using the WEIGHTING keyword, with a predefined weight function PS1997 for $w(r)$ according with the weighting form suggested by Partridge and Schwenke (1997):

$w(r) = \tanh(-\beta [(V(r) - V_{\rm min}) - V_{\rm top}] + 1.000020000200002 )/2.000020000200002$

where $\beta$ and $V_{\rm top}$ are weighting parameters and $V(r)-V_{\rm min}$ is the potential function of the corresponding field, shifted to zero.

For example, a weighting function of a potential field can be given by

abinitio poten 1 name "X 2Pi"
lambda 1
mult   2
type  grid
Weighting PS1997 1e-3 20000.0
fit_factor  1e-8
values
0.6   610516.16994  0
0.7   294361.15182  0

Here $\beta = 0.001 \frac{1}{\AA}$ and $V_{\rm top} = 20000$ cm^-1.

The Weighting feature can be used for couplings as well. In this case the values of $(V(r) - V_{\rm min})$ are taken from the potential that corresponds to the coupling in question. Here is an example for a diagonal spin-orbit field:

abinitio spin-orbit-x  A A
name "<A2Pi|LSZ|A2Pi>"
spin   0.5 0.5
lambda  1  1
sigma  0.5 0.5
factor    -i 1.175
fit_factor 1e-2
weighting ps1997  0.0001    45000.0
type  grid
<x|Lz|y>  -i -i
values
   1.58          163.05
   1.59          163.82
   ....
end

For a non-diagonal coupling between $i$ and $j$ states, Duo will use the potential corresponding to the first index $i$ to define $(V(r) - V_{\rm min})$ to calculate the weights according with the equation above.

Duo Fitting¶

Fitting ranges¶

Keywords¶

Structure of the fitting output¶

Fitting grid points¶

Constrained fit to ab initio¶

Ab initio weights¶

Example: Refinement of the BeH PEC curve¶

Table of Contents

Previous topic

Next topic

This Page