A surrogate-assisted framework for the subject-specific scalable design of compliant prosthetic wrists

This study presents a surrogate-assisted computational framework for the scalable design of compliant prosthetic wrist joints. The framework integrates computer-aided design (CAD) parameterization, automated finite element analysis (FEA), machine-learning-based surrogate modeling, and multi-objective optimization. It was applied to a compliant rolling-contact joint (CRCJ)-based prosthetic wrist capable of flexion-extension and radial-ulnar deviation. Range of motion (RoM), initial stiffness, and final stiffness were defined as design objectives to achieve biomechanically relevant wrist behavior. For each motion direction, the framework automatically generated Sobol-sequence-based CAD design variations, assigned finite element boundary conditions, and performed quasi-static FEA. This automated process produced 1024 FEA samples per direction, resulting in 4096 direction-specific datasets for surrogate-model training. Hyperparameters were optimized using Optuna, and the selected surrogate models were integrated with U-NSGA-III optimization using normalized objective errors. Five anthropometry-scaled CRCJ wrist designs were obtained from the nondominated solution set. Independent FEA re-evaluation supported the computational accuracy of the selected optimized designs, and range-of-motion testing of the fabricated XL PETG prototype showed close agreement with the design targets, achieving 84.88 deg flexion, 80.98 deg extension, 35.69 deg radial deviation, and 39.99 deg ulnar deviation.