Pareto-guided active learning for accelerating surrogate-assisted multi-objective optimization of arch dam shape

Shape optimization is one of the most critical phases in arch dam design and construction, aiming to reduce concrete volume and improve the dam stress distribution. Currently, the mainstream methods are still manual and empirical, thus lacking of efficiency and generalizability. Surrogate-assisted optimization demonstrates to be useful for enhancing structure design efficiency, yet it requires a significant amount of computationally expensive training data to ensure accurate outcomes. To accelerate the procedure, we propose a novel Paretoguided Active Learning (PgAL) framework. In the preprocessing step, the optimization mathematical model is established based on domain knowledge, and we introduce an automatic modeling technique to reduce the time cost of Finite Element (FE) simulation. Subsequently, the Gaussian Process-based PgAL is developed to accelerate the NSGA-II with the guidance of the prior information of the Pareto front. A planned ultra-high arch dam was selected as a case study, the proposed PgAL improves significantly over the traditional surrogate-assisted optimization methods, saving 70 % of the time cost to achieve similar accuracy. After optimization, the volume of dam and the volume of tensile stress region are reduced by 13.79 % and 26.57 %, respectively, achieving a good balance between economy and safety. This research provides an advanced manner for arch dam shape optimization, significantly enhancing the dam design, and may serve as a valuable reference for other similar shape optimization problems.