Abstract: In order to improve the aerodynamic performance of a turbine exhaust hood，an integrated optimization platform based on NUMECA and Isight was developed，and a multivariable optimization method was also explored. After parameterizing the diffusing structure profile of exhaust hood，the optimal Latin hypercube design of experiment was used to obtain an evenly distributed sample space. Besides，the aerodynamic performance of exhaust hood design candidate was evaluated by numerical simulation，after that，radial basis function (RBF) neural networks surrogate model was established. Two optimization methods were used to search for the final optimal resolutions，including Adaptive Simulated Algorithms (ASA) which was a global search approach and Hooke-Jeeves direct search approach. The results show that the aerodynamic performance of optimal exhaust hood is much better than that of original one on design condition without increasing the inlet static pressure，and the total pressure loss coefficient declines by 9.82%，as well as an increase by 12.2% on static pressure recovery coefficient. On the meanwhile，the velocity distribution of outlet becomes more uniform. All of these prove the effectiveness of the optimization system.

Key words: Exhaust hood；Parameterization；Design of experiment；Radial basis function ；Aerodynamic optimization

摘要： 为了改善涡轮排气蜗壳的气动性能，通过搭建NUCEMCA和Isight的集成优化平台，探索了一种多变量的优化方法，即对蜗壳参数化后，用优化的拉丁方试验设计获得空间均匀分布的试验样本，通过数值模拟求出各样本点性能参数，建立径向基神经网络代理模型(RBF)，再运用自适应模拟退火算法(ASA)和直接搜索法(Hooke-Jeeves)得到最终的优化设计参数组合。结果表明，在不提高涡轮出口静压的前提下，在设计工况，优化后的蜗壳总压损失系数在原始蜗壳的基础上降低了9.82%，静压恢复系数最高提高了12.2%，出口速度分布也更加均匀，表明了优化系统的有效性。

关键词: 排气蜗壳；参数化；试验设计；径向基神经网络；气动优化