反向
计算机科学
数学优化
可逆矩阵
人工神经网络
算法
双射
叶轮
帧(网络)
工作流程
数学
构造(python库)
反问题
采样(信号处理)
多目标优化
匹配(统计)
空格(标点符号)
控制理论(社会学)
帕累托原理
优化设计
交叉口(航空)
秩(图论)
广义逆
标识
DOI:10.1115/fedsm2026-183982
摘要
Abstract Traditional forward and inverse design approaches for centrifugal impellers both exhibit inherent limitations. In forward design, such as surrogate-based multi-objective optimization, a Pareto front can be obtained. However, each optimal solution corresponds to a specific geometry, making it difficult to capture geometric multimodality in the design space or provide meaningful uncertainty quantification. Reverse design, on the other hand, relies on experience to specify performance targets. In multi-objective problems, conflicts between objectives may lead to unrealistic or physically unattainable targets, ultimately yielding infeasible geometries. To overcome these limitations, this study develops a bidirectional design framework based on an Invertible Neural Network (INN). The INN establishes a bijective mapping between geometric parameters and a joint space of performance metrics and latent variables, ensuring exact reversibility between forward and inverse design. Within the proposed framework, the forward INN is first coupled with the NSGA-II algorithm to construct the Pareto front. Representative Pareto-optimal performance targets are then selected and supplied to the INN inverse mapping, where latent-space sampling enables generation of diverse feasible geometries. Rather than treating forward and inverse design as isolated tasks, the proposed method forms a consistent closed-loop design workflow enabled by the reversible structure of the INN. The effectiveness of the framework is demonstrated through a multi-objective design study of a centrifugal impeller.
科研通智能强力驱动
Strongly Powered by AbleSci AI