Predicting co-liquefaction bio-oil of sewage sludge and algal biomass via machine learning with experimental optimization: Focus on yield, nitrogen content, and energy recovery rate

Machine learning (ML), a powerful artificial intelligence tool, can effectively assist and guide the production of bio-oil from hydrothermal liquefaction (HTL) of wet biomass. However, for hydrothermal co-liquefaction (co-HTL), there is a considerable lack of application of experimentally verified ML. In this work, two representative wet biomasses, sewage sludge and algal biomass, were selected for co-HTL. The Gradient Boosting Regression (GBR) and Random Forest (RF) algorithms were employed for regression and feature analyses on yield (Yield_oil, %), nitrogen content (N_oil, %), and energy recovery rate (ER_oil, %) of bio-oil. The single-task results revealed that temperature (T, °C) was the most significant factor. Yield_oil and ER_oil reached their maximum values around 350 °C, while that of N_oil was around 280 °C. The multi-task results indicated that the GBR-ML model of the dataset#4 (n_estimators = 40, and max_depth = 7,) owed the highest average test R² (0.84), which was suitable for developing a prediction application. Subsequently, through experimental validation with actual biomass, the best GBR multi-task ML model (T ≥ 300 °C, Yield_oil error < 11.75 %, N_oil error < 2.40 %, and ER_oil error < 9.97 %) based on the dataset#6 was obtained for HTL/co-HTL. With these steps, we developed an application for predicting the multi-object of bio-oil, which is scarcely reported in co-hydrothermal liquefaction studies.