Abstract
Accurate prediction of bio-oil yield from pyrolysis of organic solid waste is still the main problem in the field of biomass valorization, mostly because of high variability in the composition and the heterogeneity of experimental conditions reported in the literature. Machine-learning techniques described in existing literature frequently depend on datasets that are too narrow, have limited feature representations, or shallow models that cannot sufficiently capture nonlinear thermochemical interactions that control devolatilization and liquid formation. In this study, the present study develops a deep learning-based predictive framework that addresses these gaps by using a harmonized dataset of 245 samples derived from diverse biomass sources and pyrolysis conditions. The model employs a chemically guided feature-engineering strategy that uses elemental ratios, ash-corrected volatility, and an energy-density index, followed by Variance Inflation Factor (VIF)-driven feature selection to lessen multicollinearity while keeping the mechanistic relevance. The resulting Hybrid DNPO achieves an R(2) of 0.980 and an RMSE of 1.14 on new data, making it better than all benchmark regression models, including Light Gradient Boosting (LGB). The results show that a thermochemically grounded, chemically informed deep neural prediction framework is proposed, integrating feedstock compositional descriptors and operating conditions to predict bio-oil yield from organic solid waste pyrolysis, significantly improving the performance of prior models from the literature, and can serve as a reliable tool for guiding experimental design, biomass screening, and process optimization in eco-friendly bio-oil production. By leveraging a larger, diverse dataset and advanced feature engineering, the methodology also offers a solid foundation for future studies aiming to enhance the accuracy and robustness of biomass-to-biofuel conversion predictions.