Hybrid Modeling with Artificial Neural Networks for Predicting In-Situ Bioremediation Dynamics of Diesel Fuel-Spiked Soil

Abstract

Long-term monitoring and modeling of in-situ soil bioremediation studies have their inherent challenges. In this work, the removal of diesel fuel (DF) from DF-spiked soil was studied for 138 days in six microcosm experiments, with different initial Carbon-to-Nitrogen ratios (C/N) (120, 180), and moisture content (MC) between 8 and 15% (w/w). A hybrid model predicting DF removal dynamics was proposed, where the instantaneous removal rate was modeled as an artificial neural network (ANN) function of initial C/N, MC, DF concentration, and time. DF removal rate was estimated from 250 interpolated (Akima method) points (in each experimental set) used to train the ANN model. A double-hidden layer (4–10–7–1) architecture offered the best fitness on the test subset (R²_test: 0.996), as well as on the entire dataset (R²: 0.995). LIME and SHAP analysis suggested the significance of DF concentration and MC on the ANN model explanation. Numerical integration of ANN embedded rate expression for DF removal reveals an excellent fit (R² > 0.99) to microcosm dynamics. The modeling strategy adopted in this study can be replicated in other complex bioprocess systems with limited data availability.