Comprehensive Analysis and Machine Learning Modeling of Hydrochar Characteristics: Optimizing Production Variables and Predictive Insights

Hydrothermal carbonization (HTC) at temperatures of 150-250°C and self-generated pressures of 1−5 MPa is an efficient technology to convert wet biomass wastes into stable carbon-rich solids (hydrochars) with great potential for energy production/storage, soil amendment, sustainable construction, adsorption, catalysis, etc., within a circular economy framework.

The different composition of the biomass residues, the impact of the operating conditions and the diversity of the reactions that take place during the HTC process make the design of hydrochar production very challenging.

In this study, machine learning (ML) techniques enabled addressing the inherent complexity of interactions among diverse variables and accurately modeling their relationships in order to guide the hidrochar production and properties. Three algorithms, namely Multiple Linear Regression (MLR), Support Vector Machine (SVM), and Random Forest (RF), have been chosen and systematically compared in their ability to predict the variables that exert the most influence on the development of a stable hydrochar.

Analyses were accomplished on 93 representative samples of hydrochars derived from 12 radically different bio-wastes (out-of-use woods, apple bagasse, organic fraction from municipal solid waste, sewage sludge, digestate, etc.) subjected to HTC at 180, 200 and 230 °C for 2 and 4 h.

The modeling of four key performance indicators associated with production and quality of hydrochar, such as H/C and O/C atomic ratios, the calorific value (expressed as higher heating value, HHV) and yield have been developed. The approach relies on the comprehensive analysis of a number of dependent variables categorized into six main groups: set-up parameters, operational parameters, hydrochar characterization (proximate and ultimate analysis) and thermal properties under inert (N₂), oxidative (air) and reactive (CO₂) atmospheres. Each category addresses specific aspects of the HTC process and/or hydrochar formation and its properties.

The results show that SVM achieves a better goodness of fit for H/C (R²=0.88), while RF for O/C (R²=0.92), HHV (R²=0.96), and yield (R² = 0.88) variables, both of them no-parametric algorithms. Regarding the dependent variables, the most influential categories in predicting H/C are those associated with hydrochar characterization and combustion thermogram parameters, being the variable with the greatest importance, the fixed carbon, associated to the solid carbon that remains in char after devolatilization. For O/C, those related to hydrochar characteristics and pyrolysis thermogram parameters have a relevant role. The HHV is determined by parameters of hydrochar characterization and gasification thermograms, being the most important variables the fixed carbon, associated to the solid carbon that remains in char after devolatilization, and the reactivity when 1000 °C reached and 30 and 60 minutes passed. The results obtained for yield indicate that the most important category is operational parameters, being the variables with the greatest significance, the energy, indicative of energetical harvesting potential, and densification ratio, indicative of energetical improvement.

This investigation belongs to the European Union's Horizon 2020 research and innovation program, under grant agreement No. 101006656 (GICO Project), and also to the Agroalimentación Cero Emisiones project funded by Misiones Científicas del Principado de Asturias 2022 AYUD/2022/24227 (Spain).