Modeling supercritical water gasification of municipal waste: Machine learning and data augmentation approaches

Abstract

This study presents a comprehensive machine learning framework for modeling supercritical water gasification (SCWG) of organic fraction of municipal solid waste (OFMSW) to optimize syngas production. A systematic experimental campaign investigated 16 conditions across 400–450 °C, 10–60 min reaction times, and OFMSW/H₂O ratios of 0.050–0.085, yielding gas compositions with H₂ concentrations of 12–44 %, CO₂ of 24–79 %, CO of 2.8–15.7 %, and CH₄ of 0–22 %. To address limited experimental data challenges, a SCWG-tailored Monte Carlo augmentation with compositional closure, measurement-uncertainty modeling, and leakage-free cross-validation was implemented to address tiny-n multi-output composition prediction, incorporating experimentally characterized uncertainties and generating augmented datasets of 10 × and 100 × the original size while enforcing physical feasibility (non-negativity, unit-sum, and range bounds). Random Forest (RF) and Gradient Boosting (GB) algorithms were employed for multi-output modeling of gas yield and composition. Under physically constrained augmentation and cross-validation, high in-domain accuracy is achieved across the explored operating window with R² values increasing from 0.445 ± 0.327 (RF) and 0.549 ± 0.324 (GB) on original data to > 0.96 for major components on 100 × augmented datasets. Process optimization identified distinct optimal conditions: hydrogen fuel production (450 °C, 16 min, yielding 39.4 % H₂ with 77.6 % gas yield) and Fischer-Tropsch synthesis (450 °C, 10 min, achieving H₂/CO ratio of 2.13 with 63.6 % gas yield). The framework successfully bridges limited experimental data and reliable process optimization, providing validated methodology for advancing SCWG technology toward industrial implementation. Independent validation on new waste batches and configurations will be required to establish broader generalization.