IOP PUBLISHING METROLOGIA
Metrologia 44 (2007) L23–L30 doi:10.1088/0026-1394/44/3/N02
SHORT COMMUNICATION
Evaluating the uncertainties of data
rendered by computational models
Raul R Cordero
1
, Gunther Seckmeyer
1
and Fernando Labbe
2
1
Leibniz Universit¨ at Hannover, Herrenh¨ auser Str. 2, 30419 Hannover, Germany
2
Universidad T´ ecnica Federico Santa Maria, Ave. Espa ˜ na 1680, Valpara´ ıso, Chile
Received 30 January 2007
Published 21 May 2007
Online at stacks.iop.org/Met/44/L23
Abstract
Computational models allow calculation of the value of an output quantity
from a set of linked input quantities. The value of the output quantity
yielded by a model is evidently influenced by errors in the determination of
the input quantities. Therefore, the uncertainties of the output data can be
expressed in terms of the uncertainties of the input quantities by using a
Monte Carlo-based uncertainty propagation technique. As an example, we
evaluated the uncertainty of the spectral UV irradiance rendered by a
radiative transfer model under cloudless sky conditions. This model allows
calculation of the spectrally resolved solar UV irradiance from some set of
measured input quantities linked with the concentration of atmospheric
constituents, the surface reflectivity as well as the spectral characteristics of
the aerosol modulation. Although only a single model was used in this
work, the methodology applied to evaluate the uncertainty is general and can
be applied to any other computational model.
1. Introduction
By international accord, the standard uncertainty associated
with the best estimate of a measurand can be calculated from
the standard deviation of the probability density function (PDF)
that describes the state of knowledge about the measurand.
The PDF of a non-varying stable quantity can be inferred
by measuring directly and repeatedly under repeatability
conditions the measurand with an instrument that shows
negligible drift during the observation period [1, 2].
Frequently the measurand is not directly measured but is
calculated through a measurement model that allows an output
quantity (the value of the measurand) to be expressed in terms
of linked input quantities (usually directly measured primary
quantities). In this case, the standard uncertainty of the output
quantity can be obtained from the standard uncertainties of the
input quantities by using the well-known law of propagation
of uncertainties (LPU) [2]. However, the LPU renders reliable
standard uncertainties only if the involved measurement model
is linear or weakly non-linear; otherwise, a general procedure
based on a Monte Carlo simulation, known as the law of
propagation of distributions (LPD) [3, 4], should be applied.
The LPD is valid for both linear and non-linear models and
yields the PDF of the output quantity, whose standard deviation
can be in turn used to evaluate the standard uncertainty.
The data rendered by a computational model depend
on the values of the quantities needed to run the model.
Then, if the calculated value corresponds to a measurand,
the computational model becomes a measurement model that
allows the output quantity (the value rendered by the model)
to be evaluated from the values of the input quantities (those
needed to run the model). Because a computational model can
involve a set of strongly nonlinear equations, we argue that the
uncertainty of the calculated values rendered by the model can
be evaluated by applying the LPD.
In this paper, we report on the application of a Monte
Carlo-based uncertainty propagation technique to evaluate the
uncertainties of the data rendered by a computational model.
The uncertainty evaluation involves calculating the output of
the selected model a large number of times by using sets of
input data generated according to the PDFs attributed to the
input quantities needed to run the model. Then, the mean
and standard deviation of the outputs, obtained by the large
number of calculations, are numerically computed and taken
to be the best estimate and its associated standard uncertainty,
respectively.
0026-1394/07/030023+08$30.00 © 2007 BIPM and IOP Publishing Ltd Printed in the UK L23