A Deep-Mining and Efficient Architecture Integrating Generative Modeling, Adaptive Optimization, and Multi-Task Prediction for Multi-Objective Solvent Design

Abstract Efficient solvent design is crucial for advancing sustainable separation processes, yet current approaches often suffer from narrow exploration of chemical space, limited efficiency, and difficulties in balancing competing objectives. In this study, we propose a deep learning-based molecular design architecture that integrates a molecular graph variational autoencoder for structure generation, multi-task intelligent models for property evaluation, and adaptive optimization for candidate search. Initially, the deep generative model is trained with a diverse molecular structure library for efficient molecular design. Key performance indicators, such as separation efficiency and environmental impacts, are systematically evaluated for generated molecules using multi-task predictive models and integrated into a fitness function to guide the optimization of molecular structures. The adaptive optimization algorithm iteratively refines the search direction, forming an iterative workflow to identify candidate solvents that exhibit targeted performance. The proposed integrated architecture successfully identifies six novel solvents for the extractive distillation of an azeotropic mixture benzene/n-hexane. It effectively resolves the trade-offs among molecular validity, diversity, and multiple objectives, while enhancing design efficiency and interpretability. This study offers an interpretable and computationally efficient tool for molecular design toward an intelligent and integrated paradigm, contributing to the sustainable advancement of chemical separation processes.