Machine Learning-Based Design Optimization for Uncertain Rotor Systems

This paper proposes a multi-objective stochastic optimization method that integrates a meta-heuristic algorithm with a Radial Basis Function Neural Network (RBFNN) to address rotor system design under high-dimensional uncertainties. First, a regularized RBFNN model based on Tikhonov theory is trained using Latin hypercube sampling and finite element simulations to efficiently predict dynamic responses. Then, the uncertainty optimization problem was reformulated as a bi-objective deterministic model using mean-variance optimization and chance constraints, which was further simplified into a single-objective problem via linear weighting and penalty functions. Finally, the method is applied to optimize the critical speed margin and vibration amplitude of a supercritical CO₂ rotor. Results show that RBFNN enhances global search capability, improves optimization efficiency, and ensures structural reliability under uncertainty.