Prediction of formaldehyde emissions from wood-based panels containing amino plastic resins using non-linear regression analysis and machine learning

Determining formaldehyde emissions is a key component of source control for wood-based panels. The non-linear WKI_2 equation, a modified form of the Andersen equation, is frequently used to deduce material-related emission parameters from measured test chamber concentrations and boundary conditions such as temperature (T), relative humidity (RH), air exchange (AC), and loading (L). The coefficients of the WKI_2 equation are determined using non-linear least-squares algorithms to ensure that the calculated steady-state concentrations correspond as closely as possible to the experimental data. Since the WKI_2 equation had previously only been verified for particleboard, the question arose whether it could also be used to describe the formaldehyde emissions of other wood-based panels. Therefore, a supplementary experimental dataset with medium-density fiberboard and plywood was provided and the statistical analysis showed very good agreement with the dataset for particleboard. Non-linear regression analysis is often sophisticated, as the six coefficients of the WKI_2 equation are partially correlated and the fitting procedure can be sensitive to the initial values. An alternative approach is neural network regression. In general, the advantage of machine learning lies in the recognition of non-linear patterns, which are difficult to represent in classical models without explicit physical assumptions, and in its greater robustness against measurement errors. However, neural networks require training to identify the optimal number/combination of hidden layers and neurons for the specific problem. The comparison conducted in this work has shown that both approaches, model function and machine learning, offer great potential for more precise and reliable emission predictions.