Bayesian analysis for inversion of aerosol size distribution data

Gurumurthy Ramachandran, Milind Kandlikar

Research output: Contribution to journalArticle

Abstract

Obtaining the continuous aerosol size distribution from a set of discrete measurements is an ill-posed problem. We use a new methodology based on a synthesis of Bayesian probability analysis and Monte Carlo simulations to estimate the parameters of a bimodal lognormal size distribution, i.e. the mass median diameters and geometric standard deviations of the two modes, and the fraction of the total mass in the first mode. In the Bayesian view, a measurement process serves to refine previous knowledge of physical parameters by narrowing their probability distributions. The approach avoids the direct solution of the inverse problem; instead the 'forward' problem is solved repeatedly for many sets of input size distribution parameters. From a statistical analysis of the model outputs, the input size distribution parameters are inferred. The method makes explicit use of information available prior to making any measurements and the nature of the measurement errors. Once the updated probability distributions of the size distribution parameters are determined, a number of possible solutions can be obtained. The choice could either be the maximum likelihood solution, or a minimum variance Bayes estimate that minimizes the difference between the actual measurements and those calculated using the solution. Good retrievals were obtained for numerically generated data as well as experimental data. Bayes-Monte Carlo gives reconstructions that are better than those obtained using a weighted least-squares method, even though they may sometimes come with higher computation costs. The method is not limited to determining the parameters of a certain functional form of the size distribution; it is flexible enough to allow estimation of size distributions without predetermined functional forms.

Original languageEnglish (US)
Pages (from-to)1099-1112
Number of pages14
JournalJournal of Aerosol Science
Volume27
Issue number7
DOIs
StatePublished - Oct 1996
Externally publishedYes

Fingerprint

Bayesian analysis
Aerosols
particle size
Probability distributions
least squares method
inverse problem
inversion
Measurement errors
Inverse problems
Maximum likelihood
parameter
Statistical methods
statistical analysis
methodology
cost
simulation
Costs

ASJC Scopus subject areas

  • Environmental Chemistry
  • Materials Science(all)
  • Pollution

Cite this

Bayesian analysis for inversion of aerosol size distribution data. / Ramachandran, Gurumurthy; Kandlikar, Milind.

In: Journal of Aerosol Science, Vol. 27, No. 7, 10.1996, p. 1099-1112.

Research output: Contribution to journalArticle

@article{898ab532a51d49a08d1c63d6bb95cace,
title = "Bayesian analysis for inversion of aerosol size distribution data",
abstract = "Obtaining the continuous aerosol size distribution from a set of discrete measurements is an ill-posed problem. We use a new methodology based on a synthesis of Bayesian probability analysis and Monte Carlo simulations to estimate the parameters of a bimodal lognormal size distribution, i.e. the mass median diameters and geometric standard deviations of the two modes, and the fraction of the total mass in the first mode. In the Bayesian view, a measurement process serves to refine previous knowledge of physical parameters by narrowing their probability distributions. The approach avoids the direct solution of the inverse problem; instead the 'forward' problem is solved repeatedly for many sets of input size distribution parameters. From a statistical analysis of the model outputs, the input size distribution parameters are inferred. The method makes explicit use of information available prior to making any measurements and the nature of the measurement errors. Once the updated probability distributions of the size distribution parameters are determined, a number of possible solutions can be obtained. The choice could either be the maximum likelihood solution, or a minimum variance Bayes estimate that minimizes the difference between the actual measurements and those calculated using the solution. Good retrievals were obtained for numerically generated data as well as experimental data. Bayes-Monte Carlo gives reconstructions that are better than those obtained using a weighted least-squares method, even though they may sometimes come with higher computation costs. The method is not limited to determining the parameters of a certain functional form of the size distribution; it is flexible enough to allow estimation of size distributions without predetermined functional forms.",
author = "Gurumurthy Ramachandran and Milind Kandlikar",
year = "1996",
month = "10",
doi = "10.1016/0021-8502(96)00005-5",
language = "English (US)",
volume = "27",
pages = "1099--1112",
journal = "Journal of Aerosol Science",
issn = "0021-8502",
publisher = "Elsevier Limited",
number = "7",

}

TY - JOUR

T1 - Bayesian analysis for inversion of aerosol size distribution data

AU - Ramachandran, Gurumurthy

AU - Kandlikar, Milind

PY - 1996/10

Y1 - 1996/10

N2 - Obtaining the continuous aerosol size distribution from a set of discrete measurements is an ill-posed problem. We use a new methodology based on a synthesis of Bayesian probability analysis and Monte Carlo simulations to estimate the parameters of a bimodal lognormal size distribution, i.e. the mass median diameters and geometric standard deviations of the two modes, and the fraction of the total mass in the first mode. In the Bayesian view, a measurement process serves to refine previous knowledge of physical parameters by narrowing their probability distributions. The approach avoids the direct solution of the inverse problem; instead the 'forward' problem is solved repeatedly for many sets of input size distribution parameters. From a statistical analysis of the model outputs, the input size distribution parameters are inferred. The method makes explicit use of information available prior to making any measurements and the nature of the measurement errors. Once the updated probability distributions of the size distribution parameters are determined, a number of possible solutions can be obtained. The choice could either be the maximum likelihood solution, or a minimum variance Bayes estimate that minimizes the difference between the actual measurements and those calculated using the solution. Good retrievals were obtained for numerically generated data as well as experimental data. Bayes-Monte Carlo gives reconstructions that are better than those obtained using a weighted least-squares method, even though they may sometimes come with higher computation costs. The method is not limited to determining the parameters of a certain functional form of the size distribution; it is flexible enough to allow estimation of size distributions without predetermined functional forms.

AB - Obtaining the continuous aerosol size distribution from a set of discrete measurements is an ill-posed problem. We use a new methodology based on a synthesis of Bayesian probability analysis and Monte Carlo simulations to estimate the parameters of a bimodal lognormal size distribution, i.e. the mass median diameters and geometric standard deviations of the two modes, and the fraction of the total mass in the first mode. In the Bayesian view, a measurement process serves to refine previous knowledge of physical parameters by narrowing their probability distributions. The approach avoids the direct solution of the inverse problem; instead the 'forward' problem is solved repeatedly for many sets of input size distribution parameters. From a statistical analysis of the model outputs, the input size distribution parameters are inferred. The method makes explicit use of information available prior to making any measurements and the nature of the measurement errors. Once the updated probability distributions of the size distribution parameters are determined, a number of possible solutions can be obtained. The choice could either be the maximum likelihood solution, or a minimum variance Bayes estimate that minimizes the difference between the actual measurements and those calculated using the solution. Good retrievals were obtained for numerically generated data as well as experimental data. Bayes-Monte Carlo gives reconstructions that are better than those obtained using a weighted least-squares method, even though they may sometimes come with higher computation costs. The method is not limited to determining the parameters of a certain functional form of the size distribution; it is flexible enough to allow estimation of size distributions without predetermined functional forms.

UR - http://www.scopus.com/inward/record.url?scp=0030272369&partnerID=8YFLogxK

UR - http://www.scopus.com/inward/citedby.url?scp=0030272369&partnerID=8YFLogxK

U2 - 10.1016/0021-8502(96)00005-5

DO - 10.1016/0021-8502(96)00005-5

M3 - Article

AN - SCOPUS:0030272369

VL - 27

SP - 1099

EP - 1112

JO - Journal of Aerosol Science

JF - Journal of Aerosol Science

SN - 0021-8502

IS - 7

ER -