A Deep Learning-based Framework Towards inverse Green Solvent Design for Extractive Distillation with Multi-index Constraints

Despite the popularity and efficiency of group contribution methods in computer-aided molecular design applications, their accuracy is sometimes limited due to the inability to discriminate some stereoisomers and complex compounds. One of the most urgent challenges in the chemical separation field is how to design desirable green solvents accurately and rapidly. In this contribution, a framework considering sustainability and techno-economic performance is proposed to purposively and effectively design green extractive solvents. First, a deep hierarchical molecular generative model is employed to navigate uncharted territories of the chemical space for extractive solvent inverse design. Then, the improved deep learning-based predictive models are proposed to accurately and rapidly predict the V_COSMO and σ-profile, which are prerequisites for the COSMO-SAC model to calculate the infinite dilution activity coefficients of the generated molecules. Subsequently, the developed deep hierarchical molecular generative model and the improved deep learning-based predictive model are coupled with the computer-aided molecular design approach. The green extractive solvent design process is decomposed into three sub-problems with multi-index constraints. Finally, the developed framework for designing green extractive solvent is employed in an industrial case of cyclohexane/benzene mixtures separation. As a result, the 5-methyl furfural survives from 2726 generated molecules and could be considered one of the best desirable green extractive solvents for separating cyclohexane and benzene by extraction distillation. Furthermore, the proposed framework can also be applied to other separation processes, such as liquid-liquid extraction, gas absorption, and crystallization.