摘要声振耦合系统微结构拓扑优化通常通过响应分析、灵敏度计算与设计变量更新的循环迭代，最终得到收敛的优化结构拓扑构型。此优化过程存在计算成本高、效率低等问题，为此本文提出了一种基于长短期记忆(Long-Short Term Memory, LSTM)神经网络的声振耦合系统微结构拓扑优化方法。该方法的核心思想是将声振耦合系统微结构拓扑优化过程视作构型连续变化的时序信息，利用LSTM网络强大的时序信息处理能力学习构型演化的规律，最终实现基于LSTM网络的微结构拓扑优化。论文利用基于有限元-边界元法分析的微结构优化方法生成数据集，通过测试不同网络层数、单元数和时间序列输入长度确定数值性能最优的LSTM网络，最终利用LSTM网络实现对常规声振耦合系统微结构拓扑优化的全流程替代。数值算例表明，该方法在保证优化质量的前提下显著降低了计算成本，对于不同激励频率以及不同体积约束的工况均有较好的优化效果，体现了较强的泛化能力。
Abstract：The typical microstructural topology optimization for acoustic-structure interaction system is achieved through a cyclic iteration of response analysis, sensitivity calculation, and design variable updates, suffering from expensive computation cost and low computational efficiency. To address these issues, a microstructural topology optimization method based on Long-Short Term Memory (LSTM) neural network is proposed. The core idea of this method is to treat the microstructural configurations in topology optimization process as a time series. The LSTM network, known for its powerful ability to process sequential information, is used to learn the patterns of configuration evolution. The data set is generated through microstructural topology optimization process based on the finite element-boundary element coupling analysis. Numerical examples show that the trained LSTM network can accurately predict the optimization process and reduce much computational cost compared to the conventional optimization method. In addition, the influence aspects of LSTM network structure is discussed.
2025, Vol. 46 
