A deep‐learning‐enhanced multi‐objective framework for crystallization kinetics determination via image analysis

Abstract

The optimization of crystallization kinetic parameters is frequently hindered by model variability and the prohibitive computational and experimental costs of traditional statistical analysis. The development of deep learning for image analysis provides a viable solution to this challenge. Here, we proposed a multi‐objective optimization framework leveraging deep learning‐based image analysis to expand the dimensionality of observable process variables. We determined 13 kinetic parameters simultaneously using non‐dominated sorting genetic algorithm‐III by integrating three distinct residual functions—solute concentration, particle size distribution, and agglomeration rate. While the simulation results exhibited good agreement with respect to concentration profiles and agglomeration rates, notable deviations in PSD shape were observed. These discrepancies were analyzed from the perspective of numerical stability in the simulation process, specifically the stringent constraints on flux diffusion rates required to mitigate numerical dispersion. This multi‐objective optimization approach provides a scalable pathway for the high‐throughput determination of complex crystallization kinetics.