Machine Learning Model Predicts Agricultural Waste Bio-Oil Yield with High Accuracy

Researchers from the Indian Institute of Petroleum have developed a machine learning model to predict the bio-oil yield from agricultural waste pyrolysis. The model, which utilized 387 data points from various literature sources, used proximate analysis, ultimat analysis of the agricultural feedstock, and pyrolysis reaction conditions as input features. The model was trained on eight different machine-learning algorithms, with the categorical boosting model yielding the lowest root mean squared error of 2.65 and the highest R2 value of 93.2% on the testing dataset.

Key Takeaways:

• The researchers applied machine learning techniques to develop a robust and accurate model to predict the agricultural waste bio-oil yield, with a focus on identifying the crucial parameters for maximization of bio-oil yield.
• The model utilized a total of 387 data points gathered from various literature sources, and used proximate analysis, ultimat analysis of the agricultural feedstock, and pyrolysis reaction conditions as input features.
• Eight different machine-learning algorithms were deployed for this objective, including adaptive boosting, artificial neural network, categorical boosting, decision tree with bagging, k nearest neighbor, light gradient boosting, random forest, and extreme gradient boosting.
• The categorical boost model gave the lowest root mean squared error of 2.65 out of all the models and had the greatest R2 value of 93.2% on the testing dataset.
• The SHAP analysis with a bar plot and a beeswarm plot was built to identify the input variables that had the greatest impact on bio-oil production, with hydrogen content being one of the most crucial parameters for the maximization of bio-oil yield.
• The research concluded that the developed model can be used to predict the bio-oil yield from agricultural waste pyrolysis with high accuracy.

Statistics:

• 2.65: lowest root mean squared error of all models
• 93.2%: highest R2 value of all models on the testing dataset
• 387: total number of data points gathered from various literature sources
• 8: number of different machine-learning algorithms deployed for this objective
• 1348-1259: page numbers in the International Journal of Hydrogen Energy publication

Sources:

• Data-driven Modeling of Bio-oil Yield In Agricultural Biomass Pyrolysis With Machine Learning. International Journal of Hydrogen Energy, 2025;137:1248-1259.
• Elsevier - www.elsevier.com
• International Journal of Hydrogen Energy - www.journals.elsevier.com/international-journal-of-hydrogen-energy/