Data-driven robust topology optimization using surrogate modeling, model reduction, and machine learning

This paper presents a novel data-driven framework for robust topology optimization (RTO) under load uncertainty. The proposed methodology synergistically integrates model reduction, surrogate modeling, and machine learning (ML) to efficiently solve the computationally demanding RTO problem. A Polynomial Chaos Expansion (PCE) surrogate model is employed to accurately compute the statistical moments of the stochastic compliance response required for robust optimization. To drastically reduce computational cost, a linearity-assumption-enhanced substructuring method serves as the model reduction technique. Crucially, a physics-enhanced Artificial Neural Network (ANN) is developed to predict substructure shape functions in real-time, enabling rapid online evaluations during the optimization loop. The effectiveness and superiority of the proposed data-driven approach are rigorously demonstrated through comprehensive comparisons against the full Finite Element Analysis (FEA) method, the linearity-assumption-enhanced substructuring method, and validated using Monte Carlo simulations. Results confirm that the framework achieves significant computational savings while maintaining high accuracy in large-scale robust topology design under uncertainty.

Authors

• Xuan Wang School of Computer Science and Technology, Harbin Institute of Technology Shenzhen
• Weiqi Ji

Keywords

• Robust topology optimization
• Substructure method
• Polynomial chaos expansion
• Machine learning
• Data-driven