Support vector machine regression to predict gas diffusion coefficient of biochar-amended soil

Abstract

Measurement of gas diffusion coefficient (Dp) of biochar-amended soil (BAS) under varying conditions is essential for assessing the adsorption capacity and water/gas diffusion in compacted BAS. However, there is no established equation of Dp available on this topic. Also, the factors influencing gas diffusion in BAS have not been properly studied and remain unclear. Various machine learning models were employed in this paper to learn and predict the Dp of BAS based on experimental data. Six factors (i.e., degree of compaction (DOC), biochar content (BC), soil air content (SAC), gravimetric water content (GWC), degree of saturation (DS), and porosity) are considered for testing the prediction models. The epsilon radial basis function support vector regression model showed better accuracy and predictive performance (R=0.9925) than other models and was further improved by applying the feature selection technique using the multiple linear regression and tree-based models (R=0.9937). The results reveal that SAC, DS, and porosity are the main predictor variables. The SAC proved to be the most influential predictor variable based on the estimated p-value. Furthermore, the optimal Dp was established for the various DOC and BC, which could be useful in designing engineered landfill covers. The accurate model prediction and relative importance of the predictor variables could significantly minimize the experimental work volume required to determine Dp, thereby saving time and cost.