Toward predictable hydrochar properties at scale: a critical review of design of experiments-guided machine learning in hydrothermal carbonization
TL;DR
This review synthesize and critically assess how design of experiments (DoE) and machine learning (ML) can be combined to move HTC from empirical tuning toward data-informed, property-targeted engineering of hydrochar materials.
Abstract
Hydrothermal carbonization (HTC)has emerged as a versatile platform for converting wet biomass into functional carbonaceous solids whose properties can be tuned across energy- and material-relevant applications. Under subcritical aqueous conditions, HTC proceeds through coupled dehydration, decarboxylation, polymerization, and recondensation reactions. These transformations reshape elemental composition, surface chemistry, and microstructure, enabling the production of hydrochars with tunable fixed carbon content, heating value, porosity, and surface functionality. Yet, the complexity of these reaction networks, compounded by feedstock heterogeneity and strongly nonlinear parameter interactions, continues to limit predictive control of hydrochar properties and hampers reproducible, application-driven materials design. In this review, we synthesize and critically assess how design of experiments (DoE) and machine learning (ML) can be combined to move HTC from empirical tuning toward data-informed, property-targeted engineering of hydrochar materials. We show that DoE frameworks enable statistically efficient exploration of multifactorial operating spaces and generate structured datasets that quantify the main effects and interactions of key variables, such as temperature, residence time, solid-to-liquid ratio, pressure, and catalysis. Building on these data, machine learning algorithms, including artificial neural networks, ensemble methods, boosting, and Bayesian approaches, capture high-order nonlinearities beyond classical response surface models and improve the prediction of material-critical outputs, notably mass yield, higher heating value (HHV), fixed carbon, carbon retention, and textural and chemical descriptors linked to adsorption performance and electrochemical relevance. We highlight that the DoE–ML coupling is particularly valuable for multi-objective optimization, where energy densification must be balanced against the retention of functional groups and the development of porosity, depending on whether hydrochars are targeted as solid fuels, adsorbents for water and gas treatment, or precursors for advanced carbon materials. Finally, we discuss the key bottlenecks that currently limit transferability and industrial robustness, including data quality and comparability across studies, the interpretability of predictive models, and the systematic treatment of biomass variability. We also outline methodological directions for developing more reliable hybrid DoE–ML strategies to accelerate rational design of hydrochar materials.