Optimising biochar from agricultural residues: Predicting elemental composition with machine learning

Biochar, a material whose properties are critically defined by its elemental composition, has been promoted as a sustainable way to treat various biomass wastes, including agricultural residues. However, considerable variability in these compositions across studies necessitates precise predictive techniques. This research followed the PRISMA rules for data collection and study selection, compiling data on feedstock properties and pyrolysis parameters from 38 published studies. A novel Feature-oriented Imputation method was established and employed, utilizing K-Nearest Neighbours (KNN) or Random Forest (RF) imputer to fill in missing values for features with differing characteristics. The reprocessed data were then fed into six distinct datasets and analyzed using a Gradient Boosting Regression model to predict the contents of carbon (C), hydrogen (H), oxygen (O), nitrogen (N), phosphorus (P), and potassium (K) in biochar. The rigorous machine learning process yielded excellent accuracy rates: C (R² = 0.9088, RMSE = 4.0614), H (R² = 0.9068, RMSE = 0.4180), O (R² = 0.9172, RMSE = 2.6475), N (R² = 0.8950, RMSE = 0.3416), P (R² = 0.9699, RMSE = 0.0244), and K (R² = 0.9464, RMSE = 0.3842). A comprehensive analysis of feature importance revealed that feedstock properties generally hold more significance in determining the elemental composition of biochar compared to pyrolysis parameters. The highest heating temperature (HHT) emerged as the most influential parameter for the content of H and O, while the contents of N, P, and K were predominantly determined by their respective levels in the feedstock. From these insights, optimal pyrolysis parameters were derived to tailor biochar with different elemental compositions for various applications. The developed models offer a robust framework for predicting the elemental compositions of biochar derived from various agricultural biomass, thereby eliminating the need for complex and resource-intensive laboratory trials.