4 # The contents of this file are subject to the terms of the Common Development
5 # and Distribution License Version 1.0 (the "License").
7 # You can obtain a copy of the license at
8 # http://www.opensource.org/licenses/CDDL-1.0. See the License for the
9 # specific language governing permissions and limitations under the License.
11 # When distributing Covered Code, include this CDDL HEADER in each file and
12 # include the License file in a prominent location with the name LICENSE.CDDL.
13 # If applicable, add the following below this CDDL HEADER, with the fields
14 # enclosed by brackets "[]" replaced with your own identifying information:
16 # Portions Copyright (c) [yyyy] [name of copyright owner]. All rights reserved.
22 # Copyright (c) 2013--2019, Regents of the University of Minnesota.
23 # All rights reserved.
31 This directory (utility_forces_numer_deriv) contains a Verification Check that
32 compares the forces computed by a KIM-compliant Model with those obtained using
33 numerical differentiation of the Model energies.
37 printf "model_name" | ./utility_forces_numer_deriv
41 1. It checks a single configuration: a perturbed fcc cluster made of a random
42 arrangement of the supported species with spacing scaled by a factor based
43 on the Model's cutoff value.
45 2. The code computes the DIM*N components of the force vector in two ways:
46 directly using the Model's force calculation (`force_model') and by
47 numerical differentiation using Ridders' method (`force_numer'). The
48 difference between these values in absolute value is also printed (`Force
49 diff') as well as the error predicted by Ridders' method (`pred error').
50 Ideally, one would expect the force difference to be less than the predicted
51 error. If that is the case the term `ok' is printed at the end of the
52 line. (Most potentials fail this check for most degrees of freedom, so this
53 is not a good criterion.)
55 In addition, the following normalized measure is computed:
57 \alpha = 1/(DIM*N) sqrt( \sum_i w_i (f^{model}_i - f^{numer}_i)^2 / \sum_i w_i ),
59 where w_i is the weight associated with term i:
61 w_i = 1/\hat{\epsilon}_i,
63 and \hat{\epsilon}_i is the normalized error in the numerical calculation:
65 \hat{\epsilon}_i = max(\epsilon^{numer}_i, \epsilon^{prec}) /
66 max(\force^{numer}_i, \epsilon^{prec})
68 The max functions impose a lower limit on computed values equal to the
69 numerical precision of the computer (\epsilon^{prec}).
71 \alpha has units of force and can be understood as the average error per
72 degree of freedom. The smaller the value of \alpha the more accurate the
73 force calculation of the Model.
75 In the expression for \alpha each term is weighted by the relative accuracy
76 of the numerical estimate for the derivative. Thus if f^{numer}_i is a poor
77 estimate for the derivative (i.e. \epsilon^{numer}_i is large relative to
78 f^{numer}_i) then term i contributes less to \alpha since it is less
81 In addition to \alpha, the maximum term contributing to \alpha is
82 identified. For this degree of freedom, the error and normalized error are
85 forcediff_i = |f^{model}_i - f^{numer}_i}|
87 forcediff_i/|f^{model}_i|
89 This is *not* the largest error (i.e. the largest value of forcediff_i) due
90 to the weighting term w_i. Simply looking at the maximum value of
91 forcediff_i could be misleading since the error in the numerical term,
92 \epsilon^{numer}_i, may be large for the component. Instead the maximum of
93 w_i*forcediff_i is a compromise that in some sense seeks the largest error
94 for components where the numerical derivative are also as accurate possible.