Few-shot learning to predict nitrate and phosphate adsorption capacity of biochar: A cost-effective and innovative approach

Abstract

Phosphate and nitrate contamination significantly threaten aquatic ecosystems by causing eutrophication. Biochar has gained attention as a promising adsorbent due to its high porosity and functional group availability. However, experimental characterization of the adsorption properties of biochar under varying conditions necessitates substantial time and resources. This study aimed to predict the adsorption capacities of nine distinct biochars, synthesized under varying conditions in our previous studies, for phosphate and nitrate removal. A Model-Agnostic Meta-Learning (MAML)-based few-shot learning framework was utilized to analyze a comprehensive dataset of 1350 experimental data points. Unseen biochars were excluded from the training phase to simulate real-world scenarios with limited prior data accurately. Recursive feature elimination was performed, and four features were removed to improve predictive accuracy. The findings demonstrated that using a limited set of five samples per unseen biochar enabled statistically valid and accurate predictions, as indicated by a coefficient of determination (R-2 = 0.882) and analysis of variance (p < 0.05). Furthermore, shapley additive explanations identified initial pollutant concentration, reaction time, and temperature as critical factors influencing adsorption performance. By reducing the number of experimental repetitions, the few-shot learning approach achieved a 73.46 % reduction in experimental costs. These results validate the effectiveness of the proposed framework in accurately and efficiently predicting the adsorption capacity of unseen biochars, thereby supporting cost-effective water treatment solutions. Moreover, this framework is not confined to a specific adsorbate type and shows significant potential for broader applications, including emerging contaminants such as pharmaceuticals and conventional pollutants like heavy metals.