Inverse Modeling of Pulsed Solvent Extraction Columns Using a Parameter-Screened PINN–ADM Framework

In the reprocessing of spent nuclear fuel, the coupled mass transfer between aqueous and organic phases in pulsed solvent extraction columns governs the distribution of key components. However, experimental observations in practical operations are often highly sparse in both space and time, which limits the reliability of conventional parameter estimation and modeling approaches. In this work, a unified inversion modeling framework combining physics-informed neural networks (PINNs) with the axial dispersion model (ADM) is developed to simultaneously reconstruct the concentration field and identify key model parameters. The proposed approach embeds the governing ADM equations and sparse observational data into a unified loss function, enabling the learning of continuous concentration fields without explicit mesh discretization. To address the strong parameter coupling and poor identifiability in the ADM model, a Jacobian-based identifiability analysis combined with a morphology-aware parameter screening strategy is introduced to determine the subset of parameters for inversion. Numerical results show that even under highly sparse observations covering only 1%–10% of the spatiotemporal domain, the proposed method can reliably reconstruct the concentration field with a global relative L 2 error on the order of 10 –2 . Meanwhile, the estimation accuracy of key parameters such as the distribution ratio is significantly improved compared with full-parameter inversion without screening. The proposed PINN–ADM framework provides a stable and data-efficient approach for inverse modeling of coupled mass transfer systems and has potential applications in process modeling and monitoring of solvent extraction operations.