Machine learning optimization for algal biochar yield: integrating experimental validation and sensitivity analysis

Abstract Pyrolysis requires extensive experimentation to achieve optimum thermochemical conversion, which can be addressed by integrating machine learning (ML) predictive solutions. The abundant availability of algae with low volume footprint makes it viable green biomass to achieve thermochemical products. For optimum algal biochar (BC) yield production, the relation of ultimate, proximate analysis with process conditions is critical. This study’s objective is twofold: It aims to develop a robust ML model, trained on diverse literature data and optimized using particle swarm optimization and genetic algorithm, that predicts BC yield across various feedstocks and conditions. Secondly, the optimum process parameters are derived to maximize BC yield with the experimental validation for the collected samples at their respective chemical and structural compositions. Limited data points for algal biomass induce a comparative analysis of ML models, including Gaussian process regression, ensembled tree (ET), decision tree and support vector machine. The predictive capability of ET enhanced through optimization performed exceptionally well for BC yield prediction with testing R 2 = 0.77993 and RMSE = 6.9792. 2D and 3D partial dependence plots imply that BC yield is primarily influenced by pyrolysis temperature, volatile matter, and heating rate with SHAP values of 1.2785, 0.3972, and 0.2949, respectively. Monte Carlo simulation and Sobol sensitivity analysis substantiate statistically the impact of selected features on algal BC yield. Inverse optimization of ET model suggests that the maximum BC yield production is 76.33% at a temperature of 500 °C, a heating rate of 10 °C/min, a residence time of 60 min, a N 2 flow rate of 0.5 L/min, and particle size of 1.5mm. Graphical Abstract