Multimodal Heat Source Layout Optimization based on Neural Architecture Search  
Author Jialiang Sun


Co-Author(s) Xiaohu Zheng; Wen Yao; Xiaoya Zhang; Weien Zhou; Xiaoqian Chen


Abstract Deep learning surrogate assisted heat source layout optimization (HSLO) can decrease the computational cost largely. However, it faces the design difficulty of the neural network surrogate and optimization algorithms. To address the challenges, we propose a method called multimodal heat source layout optimization based on neural architecture search (MSHLO-NAS). Neural architecture search (NAS) method combined with Feature Pyramid Network (FPN) framework is developed to realize the purpose of automatically searching for a small deep learning surrogate model for HSLO. On the basis of the searched surrogate, a multimodal neighborhood search based layout optimization algorithm (MNSLO) is proposed, which could obtain more and better layout design schemes simultaneously in single optimization. The typical two-dimensional heat conduction optimization problem is utilized to demonstrate the effectiveness of the proposed method. With the similar accuracy, NAS finds models with 80% fewer parameters, 64% fewer FLOPs and 36% faster inference time than the original one. Besides, with the assistance of deep learning surrogate by NAS, MNSLO can find multiple near optimal design schemes simultaneously to provide the design diversities for designers.


Keywords Heat source layout optimization, neural architecture search, multimodal optimization
    Article #:  DSBFI23-112
Proceedings of 2nd ISSAT International Conference on Data Science in Business, Finance and Industry
January 8-10, 2023 - Da Nang, Vietnam