Application of Sequential Approximation Optimization with Reduced Simulation Numbers for Integrated System Optimization

Simulation plays a crucial role in optimizing complex system designs. While individual subsystems may have relatively short processing times, when the entire system is integrated, the significant time and computational resources required remain a considerable challenge. On the other hand, system integration introduces effects that cannot be easily formalized, making integrated simulation indispensable. To address this challenge, this study proposes a methodology to enhance accuracy with limited data by augmenting the training dataset utilizing the characteristics of the Integration Neural Networks (INN) within the framework of Sequential Approximation Optimization (SAO). The INN is a surrogate model that integrates a deductive neural network, which handles scientifically formulable phenomena, and an inductive neural network, which addresses non-formulable phenomena. In this context, the expressive constraints of the deductive neural network serve to restrict variations between training points. Consequently, the surrogate model can attain high accuracy even with a limited quantity of training data. By leveraging this attribute, the proposed method achieves improved solution accuracy in SAO with minimal data points by adding training points surrounding the predicted optimal solution. In conventional SAO, the initial training dataset starts with 10 times the number of design variables. In contrast, this study hypothesized that optimization with fewer simulation runs could be achieved by strategically replacing random, low-quality initial training data with data selected based on intermediate learning results. Specifically, we investigated whether the total amount of training data required could be reduced by decreasing the number of initial training data points and increasing the number of data points added during the optimization process. The results of the experiments showed that stable solutions within the error range were obtained for all cases. Furthermore, it was demonstrated that a balance between data reduction and optimization stability could be successfully achieved by setting the number of initial training samples to three times the number of design variables (21 samples). As a result, the number of simulation runs required to construct the training dataset was reduced by approximately 50.2% compared to the conventional SAO method, demonstrating significant potential for shortening the design cycle in product optimization.