Experimental and Machine Learning-Based Assessment of Thermal Conductivity in Biochar-Amended Soils

Accurate characterization of soil thermal conductivity is important for understanding heat transfer in geotechnical and environmental applications. This study investigates the effect of biochar amendments on the thermal conductivity of soils under controlled laboratory conditions and develops a machine learning-based predictive framework to improve estimation accuracy. Experiments were conducted on kaolin and silty sand mixed with 10% by weight of three biochars—peach pit, corn cob, and reed—under two compaction levels, loose and dense. Thermal conductivity was measured using a TEMPOS thermal analyzer. Conventional empirical and theoretical models, including those by Kersten, Johansen, and Haigh, were evaluated but showed limited accuracy for biochar-amended soils because they could not capture the non-linear influence of soil type, moisture content, dry density, compaction, and biochar type. To address this limitation, machine learning models including regression, KNN, random forest, and artificial neural network were trained using these input variables. The results showed that non-linear machine learning models outperformed traditional methods, with KNN achieving the best performance, reaching R2 = 0.96 and RMSE = 0.073. Among the biochars tested, peach pit biochar caused the least reduction in thermal conductivity. The findings demonstrate that machine learning can provide a reliable predictive tool for soil thermal conductivity under controlled conditions and offer useful insight for soil thermal assessment in standard engineering applications.