A Novel Construction Method of Bayesian Neural Networks Based on Multi-type Engineering Knowledge

Abstract In the field of engineering, the utilization of surrogate models to replace computationally intensive simulation software has become a widely adopted approach. However, when addressing complex engineering problems, the costs of simulations can escalate significantly, making it challenging for simulation data to fulfill the training requirements of surrogate models. Recognizing that designers accumulate valuable design knowledge throughout the design process, this knowledge inherently governs the mapping rules between design parameters and performance metrics. This paper introduces a novel method for constructing surrogate models by integrating limited simulation data with engineering knowledge through Bayesian neural networks (B-DaKnow). In B-DaKnow, neural networks employ variational inference and automatic differentiation to amalgamate simulation data and engineering knowledge while optimizing weights and biases via evolutionary algorithms. The proposed methodology is validated using ten benchmark functions and three engineering cases. The experimental results demonstrate that: (1) the incorporation of diverse engineering knowledge enhances prediction accuracy in B-DaKnow to varying degrees; (2) in tackling complex engineering challenges, B-DaKnow exhibits superior performance compared to alternative algorithms; (3) B-DaKnow demonstrates commendable robustness, as evidenced by only slight fluctuations in prediction results across different problems.