ParetoGen: Generative Machine Learning Models To Push the Pareto Optimal Frontier of Functionality-Hazard Trade-offs in Per- and Polyfluoroalkyl Substances Green Alternative Designs

Faced with chemical pollution that has exceeded the planetary boundaries, endeavors must be directed toward designing green alternatives to reduce the hazards of industrial chemicals. Per- and polyfluoroalkyl substances (PFASs) are critical industrial chemicals with superior functionalities but proven hazards. Trade-offs between functionalities and hazards constitute the challenge in PFAS green substitution, known as multiparameter optimization. The essential goal of multiparameter optimization is discerning chemicals for which enhancements in one property invariably lead to degradations of others. In statistical parlance, the chemicals constitute the Pareto optimal frontiers (POFs). Based on generative machine learning algorithms, this study developed a method named ParetoGen for propelling POFs of multiobjective optimization issues in green alternative designs. Results demonstrate the ParetoGen stably and significantly pushed the POFs of functionality-hazard trade-offs in the design of green alternatives to PFASs. Compared with existing PFAS alternatives, structures generated by the ParetoGen exhibit lower hazards and enhanced functionalities. By introducing the Pareto optimization theory into the molecular design of green alternatives, this study establishes a universal theoretical foundation and evaluation methodology for green chemical substitution. The findings reveal valuable insights into the chemical structural design of PFASs, and the developed method can be employed in general chemical structural design tasks.