Machine learning model with a novel self–adjustment method: A powerful tool for predicting biomass ash fusibility and enhancing its potential applications

Abstract

Biomass ash has been extensively studied for its potential applications, owing to its high content of alkali and alkaline earth metallic species (AAEMs). These AAEMs can act as catalysts in biomass thermochemical conversion and other industrial processes. However, AAEMs can also cause slagging and agglomeration, which can significantly impact system operations. To better understand these effects, we investigated the relationship between ash melting behavior and the chemical composition of biomass ash using a machine learning (ML) model. To enhance the model's performance, we employed a self-adjustment (SA) method, which significantly improved predictive accuracy. The SA-ETR model achieved an R² value greater than 0.93, based on a dataset of 268 data points. We provided a detailed explanation of the SA-optimized ML model using Python's Shapley Additive Explanations (SHAP) library, which included global and local feature importance analysis, investigation of simultaneous effects between two features, and individual data point prediction analysis. The contents of K₂O, SiO₂, CaO, and Al₂O₃ were considered as the most significant factors affecting biomass ash's initial deformation temperature (IDT). The insights gained from this study can help investors and researchers reduce experimental complexity and improve system operation.