AI-DRIVEN PREDICTIVE MODELLING OF BIOMASS THERMAL DEGRADATION USING THERMOGRAVIMETRIC ANALYSIS DATA AND ADVANCED MACHINE LEARNING

Abstract

The paper presents a new AI-powered model that predicts biomass thermal degradation with unprecedented accuracy of R² = 0.978 with sparse thermogravimetric analysis (TGA) data, and is 6% R² higher and the RMSE is 0.42% lower than the best kernel-based models. Bayesian-optimised CatBoost ensemble models separate TG/DTG profiles, the evolution of activation energies (180-265 kJ/mol) (RMSE = 4.2 kJ/mol), and multi-stage pyrolysis kinetics using single-scan TEMP-WT LOSS triplets (without the need to use parallel heating rates) and quantify the mass transfer limitations that are important in the design of chemical reactors. SHAP interpretability indicates that the dominance of DTG gradients (28.4% importance) and temperature polynomials (15.3%) are the most important predictors, connecting machine learning with chemical reaction engineering because they can capture the physics of devolatilization rates, secondary cracking and char stabilization, which are not modelled in traditional distributed activation energy models (DAEM). The framework outperforms XGBoost (R² = 0.954), SVR (R² = 0.923), Random Forest (R² = 0.941) and ANN (R² = 0.917), making empirical thermochemical analysis predictive process systems engineering instead of 70x faster. Stage-specific fidelity is justified by Graphs; overfitting-free convergence is established hierarchy of causal features used to design the best experiments is justified by results. Industrial impacts are 1.2% bio-oil yield variability compared to 5-10% of conventional kinetics, indicating 100K+/yr revenue per 10 ton/hr post due to accurate residence time optimization. Precision modelling of thermochemical processing converting lignocellulosic waste to optimized hydrogen/ bio-oil/carbon product slats is democratized by the open-source pipeline to the agricultural economies, which in turn reduces exergy waste in a commercial biorefinery.