Hydrodesulfurization (HDS) remains the primary industrial route for sulfur removal, but increasingly heavy and complex feedstocks make deep desulfurization more challenging. This review examines the evolution of HDS from model sulfur compounds to refinery-relevant heavy feeds, including vacuum gas oils, residues, biomass-derived oils, and waste tire and plastic pyrolysis oils. The distribution and reactivity of inorganic and organic sulfur species are discussed, with particular emphasis on refractory thiophenic compounds, and the competing direct desulfurization and hydrogenation pathways. The effects of temperature, hydrogen pressure, liquid hourly space velocity, and feed sulfur content are evaluated in relation to sulfur conversion, hydrogen consumption, catalyst deactivation, and process economics. The emerging role of machine learning (ML) in HDS research is assessed. ML enables quantitative mapping of catalyst descriptors, feedstock properties, and operating conditions to catalytic performance, enhancing catalyst screening, structure–activity analysis, sulfur-content prediction, soft sensing, multi-objective optimization of operating conditions, and catalyst lifetime prediction. The review further highlights hybrid physics-aware models, explainable ML, digital twins, and adaptive learning as promising approaches for bridging laboratory-scale models with dynamic refinery operation. Unlike conventional trial-and-error approaches, these data-driven methods can reveal complex interactions among catalysts, feedstock, and process variables while reducing experimental effort and enabling real-time decision support. Some barriers remain, including data quality, training-deployment data mismatch, model transferability, interpretability, and limited long-term industrial datasets. Integrating mechanistic HDS knowledge with ML promises a pathway toward faster catalyst development, optimized operation, improved catalyst utilization, and more sustainable heavy-feed desulfurization.
