Multi-stage centrifugal pumps are frequently used in high-lift applications and consume considerable energy, but suffer from poor performance and large axial force. The rear shroud of impeller is trimmed for reducing axial thrust, but this degrades performance. This study analyzes performance degradation and optimizes performance and axial force. Experiments and simulations are conducted on different ratios of rear shroud to front shroud (λ). Total pressure losses are calculated, and flow losses are visualized using the entropy generation method. Both measured and simulated performances decrease as the rear shroud is trimmed. Designs with different λ meet the head coefficient requirement of 1.1. However, λ of 0.86 has the best efficiency of 42.7%, λ of 0.83 reaches 42.5%, λ of 0.8 shows the lowest efficiency of 39.9%. Efficiency in the middle channel improves as the rear shroud is trimmed, but this cannot offset increased losses in the impeller and rear side chamber. Entropy production is exacerbated in the axial passage between impeller and rear side chamber due to the collision between impeller-driven flow and pressure-driven backflow. When λ is reduced by 0.03, axial thrust drops by 7%. To compromise between performance and axial thrust, λ should be designed at 0.83.

To address the insufficient near-wall prediction capability of the traditional Partially Averaged Navier–Stokes (PANS) model in simulating curvature flows, a new nonlinear PANS model with near-wall correction was developed in this study. The model, referred to as the CLS PANS model, is constructed based on Craft’s nonlinear stress formulation and incorporates additional dissipation source and length-scale correction terms to enhance accuracy in curved, rotating, and separated flow fields. To evaluate its applicability and reliability, the new nonlinear PANS model was applied to three representative cases: Taylor–Couette flow, flow past a circular cylinder, and internal flow in a centrifugal pump. Numerical results were systematically compared with experimental data, Direct Numerical Simulation (DNS) results, and results from conventional Reynolds-Averaged Navier–Stokes and k-ε PANS models. The results show that the new nonlinear PANS model can accurately predict complex flow structures such as Taylor vortices and herringbone streaks with lower computational cost, demonstrating improved scale-resolving capability and near-wall performance. For flow past a circular cylinder, the predicted drag coefficient, Strouhal number, and velocity distribution in the wake agree well with experiments. In the centrifugal pump case, the model effectively captured the low-speed and separated flow regions near the blade pressure surfaces, yielding results consistent with experimental observations. Overall, the new nonlinear PANS model achieves a favorable balance between accuracy and efficiency and exhibits strong potential for simulating curvature- and rotation-dominated turbulent flows.