Data-Driven evolutionary algorithm for reactive distillation: Online surrogate model and feasible domain relaxion

• Proposed OS-NSGA-III framework with online BANN adaptive sampling and validation. • Combined convergence loss function addresses RD feasible domain challenge. • Achieved 84 % reduction in Aspen simulation while maintaining Pareto accuracy. Reactive distillation (RD) faces design and optimization bottlenecks due to complex vapor–liquid-reaction equilibria and discrete parameter constraints. This work develops an online surrogate-assisted Non-dominated Sorting Genetic Algorithm (OS-NSGA-III) integrating bootstrap-aggregated neural networks (BANN) with adaptive sampling to enable efficient multi-objective RD design. The framework introduces combined convergence loss function, effectively improved the feasible domain discontinuity issue for RD design. The proposed algorithm implements adaptive sampling strategies and full and partial validation mechanisms to overcome simulation non-convergence issues while maintaining search efficiency. The algorithms are designed so that up to 90 % of calculations are based on the surrogate model, with Aspen Plus re-calibration performed periodically to maintain predictive accuracy (Mean Relative Error < 0.2). Case studies on pressure-swing RD configurations demonstrate 84 % reduction in Aspen evaluations compared to conventional methods while maintaining Pareto front coverage accuracy. This algorithm provides a computationally efficient paradigm for complex chemical process optimization with inherent simulation convergence challenges.