By using precise analytical model data to assist numerical simulation data

constructing the high-quality training dataset for the surrogate model offers lower computational cost and enhances the efficiency of multi-objective optimization. However

existing precise analytical modeling methods often impose specific requirements on machine structures

limiting their applicability. To address this issue

the work proposes an optimization design method for motor surrogate models based on geometric similarity transfer learning. This method first utilizes the geometric similarity between machines’ physical structures to construct similar motors that are able to establish precise analytical modeling. Subsequently

a hierarchical optimization strategy is adopted by partitioning the design-variable space into high- and low-sensitivity subspaces to enhance the consistency between the transfer-learned optimization results and those obtained from the original prototype. The high sensitivity parameter space with fewer variables is optimized using a conventional surrogate model established from finite element analysis (FEA) data

while the low sensitivity parameter space with more variables is optimized based on transfer learning (TL) to train a multi-fidelity surrogate model. The proposed method breaks the topological constraints of accurate analytical models

reduces the difficulty of analytical modeling for structurally complex machines

and significantly improves optimization efficiency for high-dimensional problems through hierarchical and multi-fidelity transfer learning

achieving substantial computational savings without sacrificing accuracy. The design method was applied to the multi-objective optimization design of a built-in consequent pole permanent magnet vernier motor

and prototype experimental results validated its effectiveness.