Development of a novel physics-informed machine learning model for advanced thermochemical waste conversion

Abstract

• Various data-driven models developed and compared for indirect gasification. • Physics-informed machine learning model for predicting the outputs of gasification. • Significant improvement in carbon balance closure with 10 % physics contribution. • Model covers a wide range of feedstocks ranging from biomass-rich to plastics-rich. • Model gives insights into input feature importance for various products. A physics-informed machine learning (ML) model, which incorporates the conservation of carbon mass, was developed to predict the product gas yield and composition for indirect gasification of waste in a fluidized bed. A dataset was compiled from experimental data of an in-house reactor, encompassing a wide range of feedstocks characteristics (biomass to plastics) and process conditions, which served as input for the model. Four data-driven models were trained and evaluated, with the XGBoost model having the best predictive accuracy (RMSE = 1.1 & R 2 = 0.99) and being adapted for the physics-informed model. The optimum physics contribution was 30 % (70 % data contribution) to maintain predictive accuracy (RMSE = 2.7 & R 2 = 0.95) and improve carbon closure. Feedstock properties were shown to have a higher feature importance compared to the operating conditions. The developed physics-informed model demonstrated the potential of ML models for the modelling of gasification of various waste streams. This is a promising first step towards improving data-driven ML models for application to thermochemical systems.

Topics & Concepts