Characterizing sludge pyrolysis by machine learning: Towards sustainable bioenergy production from wastes

Sludge pyrolysis has sparked the interest of researchers because of its capability to dispose of hazardous residues while producing valuable bioproducts. Numerous expensive and laborious experiments are conducted to understand sludge pyrolysis. Machine learning technology can eliminate the need for experimental measurements by systematically learning relationships between variables from historical data. This research aimed to propose a machine learning model to characterize sludge pyrolysis products. A comprehensive database covering various sludge types and pyrolysis reaction conditions was constructed from experimental data. The k-nearest neighbor algorithm was used to reconstruct the missing inputs of sludge composition. The principal component analysis method was then used to decrease dataset dimensionality and acquire relevant information. The obtained scores were normalized and introduced into three machine learning models. The input variables were the chemical properties of sludge and reaction conditions. The response parameters were the distribution and composition of pyrolysis products. Based on descriptive data analysis, the optimum bio-oil yield was obtained at temperatures between 500 and 600 °C. At higher temperatures (700–800 °C), a transition was observed in the product distribution towards more syngas. The random forest regression model showed the highest accuracy among the applied models, with a correlation coefficient higher than 0.813 and a relative mean squared error lower than 12.51. The SHAP analysis using the random forest algorithm was successfully conducted to understand the importance of input variables on output responses. The five top significant features affecting bio-oil yield were ash content, fixed carbon content, operating temperature, and volatile matter content.