Technical note: Monte Carlo genetic algorithm (MCGA) for model analysis of multiphase chemical kinetics to determine transport and reaction rate coefficients using multiple experimental data sets

Abstract

Abstract. We present a Monte Carlo genetic algorithm (MCGA) for efficient, automated, and unbiased global optimization of model input parameters by simultaneous fitting to multiple experimental data sets. The algorithm was developed to address the inverse modelling problems associated with fitting large sets of model input parameters encountered in state-of-the-art kinetic models for heterogeneous and multiphase atmospheric chemistry. The MCGA approach utilizes a sequence of optimization methods to find and characterize the solution of an optimization problem. It addresses an issue inherent to complex models whose extensive input parameter sets may not be uniquely determined from limited input data. Such ambiguity in the derived parameter values can be reliably detected using this new set of tools, allowing users to design experiments that should be particularly useful for constraining model parameters. We show that the MCGA has been used successfully to constrain parameters such as chemical reaction rate coefficients, diffusion coefficients, and Henry's law solubility coefficients in kinetic models of gas uptake and chemical transformation of aerosol particles as well as multiphase chemistry at the atmosphere-biosphere interface. While this study focuses on the processes outlined above, the MCGA approach should be portable to any numerical process model with similar computational expense and extent of the fitting parameter space.

Optimization of process models for determining volatility distribution and viscosity of organic aerosols from isothermal particle evaporation data

“…The Monte Carlo Genetic Algorithm (MCGA) is an optimization method developed by Berkemeier et al, 2017. MCGA has been previously used in estimating atmospheric multiphase chemistry parameters such as reaction rate coefficients and bulk phase diffusion coefficients (Table 1 in Berkemeier et al, 2017).…”

“…The algorithm divides the optimization process into two different parts: a random sampling of the parameter space (MC part) and a genetic algorithm (GA part) with an initial population from the MC sampling. Random sampling means that a predetermined number of parameter sets, named candidate solutions or candidates, are created by randomly choosing values for the free parameters (Berkemeier et al, 2017). These candidates form a population.…”

“…The initial population to the GA part is chosen from the MC part so that 5% of the best-fit candidates are chosen and the remaining 95% of the population is chosen randomly. The computation times of the MC and GA parts are divided, so that both parts require about half of the total computation time (Berkemeier et al, 2017).…”

“…A new candidate is created from the two parents' parameter sets. A new parameter for the new candidate is chosen from the parents' parameters randomly (Berkemeier et al, 2017).…”

“…The version of the algorithm used in this work differs from the version described in Berkemeier et al, (2017) in that how a new candidate is accepted to the next generation and how parents are chosen. Once the new candidate has all its free parameters drawn, it is accepted to the new generation if its fitness is lower than the largest fitness in the previous generation or if the largest fitness in the previous generation divided by the candidate's fitness is lower than a uniform random number between 0 and 1.…”