Barrier-Balancing Machine Learning Framework with Multidimensional Descriptors for Interpretable Screening of Secondary Amine Ionic Liquid Catalysts

Rational screening and design of homogeneous catalysts remain a core challenge in multistep reaction systems, particularly when multiple elementary steps exhibit comparable and elevated activation barriers. Inspired by the Sabatier principle and the degree of rate control theory, we propose a barrier-balancing machine learning framework that focuses on coordinating rather than minimizing individual energy barrier with density functional theory calculations and experimental validation, to identify efficient secondary amine ionic liquid catalysts. Taking the Mannich reaction to synthesize methylacrolein as a probe, we target two kinetically significant steps of nucleophilic attack on formaldehyde (ΔE1) and decomposition of the intermediate (ΔE2), and define the maximum of the two as the model’s training target. A physically interpretable multidimensional descriptor library is constructed, integrating physicochemical-based properties, Morgan fingerprints, volume-based steric parameters, and quantum chemical features. The trained random forest model, interpreted via SHapley Additive exPlanations (SHAP), identifies key descriptors that govern barrier heights. A barrier-based scoring metric derived from model predictions shows strong correlation with experimental selectivity, validating its physical relevance and predictive power. This framework not only differentiates aryl- and alkyl-substituted amines by their electronic environments, but also offers an interpretable and generalizable tool for catalyst screening in complex homogeneous systems.