Parameter estimation and model selection for water sorption in a wood fibre material

TL;DR

This study evaluates eight models for water sorption in wood fibre materials, analyzing their parameter estimation reliability and selecting the best model using Bayesian computation, with implications for modeling heat and mass transfer in porous materials.

Contribution

It introduces a comprehensive assessment of model identifiability and employs Bayesian methods for simultaneous model selection and parameter estimation in water sorption modeling.

Findings

Thermodynamic and Feng-Xing models performed best in selection.

Seven models showed low primary identifiability for at least one parameter.

Bayesian computation effectively identified the most suitable model and parameters.

Abstract

The sorption curve is an essential feature for the modelling of heat and mass transfer in porous building materials. Several models have been proposed in the literature to represent the amount of moisture content in the material according to the water activity (or capillary pressure) level. These models are based on analytical expressions and few parameters that need to be estimated by inverse analysis. This article investigates the reliability of eight models through the accuracy of the estimated parameters. For this, experimental data for a wood fibre material are generated with special attention to the stop criterion to capture long time kinetic constants. Among five sets of measurements, the best estimate is computed. The reliability of the models is then discussed. After proving the theoretical identifiability of the unknown parameters for each model, the primary identifiability is analysed. It evaluates whether the parameters influence on the model output is sufficient to proceed the parameter estimation with accuracy. For this, a continuous derivative-based approach is adopted. Seven models have a low primary identifiability for at least one parameter. Indeed, when estimating the unknown parameters using the experimental observations, the parameters with low primary identifiability exhibit large uncertainties. Finally, an Approximation Bayesian Computation algorithm is used to simultaneously select the best model and estimate the parameters that best represent the experimental data. The thermodynamic and \textsc{Feng}--\textsc{Xing} models, together with a proposed model in this work, were the best ones selected by this algorithm.