Machine learning-driven prediction and optimization of selective glycerol electrocatalytic reduction into propanediols

Abstract

Electrochemical conversion of crude glycerol–a surplus by-product of biodiesel production–into value-added propanediols (PDO) presents a sustainable bioresource valorization. However, optimizing selective glycerol electrocatalytic reduction (ECR) remains challenging due to complex interactions among multiple reaction parameters. Here, we employ an integrated machine learning-driven optimization framework combining XGBoost with particle swarm optimization (PSO) to predict and optimize glycerol ECR performance, targeting both conversion rate (CR) and electroreduction product yields (ECR PY). A dataset of 446 experimental datapoints curated from published literature was used to train the XGBoost model, achieving high prediction accuracy (R² of 0.98 for CR; 0.80 for ECR PY), outperforming other algorithms and demonstrating robustness against unbalanced datasets. Feature analysis revealed that low-pH electrolytes and longer reaction times significantly enhance both outputs, while higher temperatures and carbon-based electrocatalysts positively influence ECR PY by facilitating CO bond cleavage in glycerol. XGBoost-PSO optimization predicted maximum CR (100 %) using a Pt cathode at 24.15 h, 24.66 °C, pH 1.08, 66.96 rpm stir rate, 0.43 M electrolyte concentration, and 0.28 A/cm² current density. Meanwhile, the highest ECR PY (53.29 %) was predicted with a carbon cathode at 22.27 h, 78.87 °C, pH 0.99, 650.18 rpm, 3.84 M electrolyte, and 0.14 A/cm². Experimental validation confirmed the model's predictive accuracy within ∼10 % error. GC–MS further validated the selective formation of PDOs, with yield of 21.01 % under optimized conditions. This framework offers a robust, data-driven alternative to traditional trial-and-error approaches, providing mechanistic insights and practical guidance for scalable, economically viable glycerol ECR in biodiesel industry.