Multi-objective optimization strategy for green solvent design via a deep generative model learned from pre-set molecule pairs

Abstract

Green solvent design is usually a multi-objective optimization problem that requires identification of a set of solvent molecules to balance multiple, often trade-off, properties. At the same time, process constraints need to be addressed since solvent properties impact the process feasibility like in the extractive distillation separation process. Hence, a green solvent multi-objective optimization framework is proposed with EH&S properties, process constraints, and energy consumption analysis, where the molecular design optimization model relies upon the ability of the proposed infinite dilution activity coefficient (IDAC) direct prediction model to accurately predict process properties in addition to molecular properties. The process properties are short-cut properties of the extractive distillation process, namely selectivity and solution capacity. To this end, the proposed IDAC direct prediction model is employed to prepare molecule pairs with selectivity and solution capacity improvement constraints to train the molecular multi-objective optimization model, which can learn the optimization path from the pre-set molecule pairs and then optimize a given solvent via the prediction of a disconnection site and molecular fragment addition or removal at that site. An extractive distillation process to separate a cyclohexane/benzene mixture is taken as an example to demonstrate the proposed framework. As a result, three candidate green solvents are optimized and designed to recover benzene from mixtures of benzene and cyclohexane. The proposed green solvent multi-objective optimization framework is flexible enough to be employed in other chemical separation processes, where solvent property assessment is needed to evaluate the feasibility and performance of the processes.