Enhancing chromatographic resolution via smoothing-based global optimization and neural surrogate PDE modeling

Chromatographic resolution is an issue of significant importance in preparative chromatography. However, due to the high-dimensional parameter space and nonlinear behavior, baseline separation in multi-component preparative chromatography remains challenging. To overcome these difficulties, we propose a surrogate-assisted optimizer focused on chromatographic resolution that uses a feedforward neural network to emulate PDE-based outlet responses. Based on this surrogate model, we employ the global optimization method of GS-PowerOpt to reshape the nonconvex landscape and robustly identify high-resolution designs. We prove the continuity of the resolution function and establish guarantees for convergence in the smoothed search. Through several case studies, the method consistently identifies designs with well-separated peaks and high resolution. The results offer a practical and theoretically grounded route to effective method development.