Sequential Monte Carlo with likelihood tempering and parallel implementation for uncertainty quantification

Abstract Bayesian estimation enables uncertainty quantification, but analytical implementation is often intractable. As an approximate approach, the Markov Chain Monte Carlo (MCMC) method is widely used, though it entails a high computational cost due to frequent evaluations of the likelihood function. As an alternative approach, the sequential Monte Carlo (SMC) method is well known for time‐series data, while its advantages for non‐time series data using likelihood tempering have not been fully explored. In this study, we implemented an SMC algorithm with likelihood tempering and compared it with MCMC through two types of problems: toy problems and parameter uncertainty quantification problems for a methanation reactor model using real experimental data. In addition, we evaluated the effect of parallelization on SMC performance. The results showed that SMC with likelihood tempering exhibits substantially better computational efficiency and stability for problems with strong parameter correlations and with likelihood evaluations that require computationally expensive simulations.