一种使用BP神经网络加速的蒙特卡洛拉格朗日水滴求解器  

A Monte Carlo Lagrangian Droplet Solver with Backpropagation Neural Network for Aircraft Icing Simulation

作　　者：刘宇 屈经国 易贤[1,2,3] 王强 LIU Yu;QU Jingguo;YI Xian;WANG Qiang(Low Speed Aerodynamic Institute,China Aerodynamics Research and Development Center,Mianyang 610082,P.R.China;Anti/De‑Icing Key Laboratory,China Aerodynamics Research and Development Center,Mianyang 610082,P.R.China;Key Laboratory of Aerodynamics,Mianyang 610082,P.R.China;School of Computer Science,Southwest Petroleum University,Chengdu 610500,P.R.China)

机构地区：[1]中国空气动力研究与发展中心低速空气动力研究所,中国绵阳610082 [2]中国空气动力研究与发展中心结冰与防除冰重点实验室,中国绵阳610082 [3]空气动力学国家重点实验室,中国绵阳610082 [4]西南石油大学计算机科学学院,中国成都610500

出　　处：《Transactions of Nanjing University of Aeronautics and Astronautics》2023年第5期566-577,共12页南京航空航天大学学报（英文版）

摘　　要：结冰可能威胁飞行安全。拉格朗日方法被广泛应用于求解结冰过程中的水收集系数,但是其发展受到鲁棒性问题和高计算成本限制。为了弥补拉格朗日方法的缺陷,使用蒙特卡洛积分法和反向传播(Backpropagation,BP)神经网络分别用于解决鲁棒性问题和降低计算成本。基于蒙特卡洛方法的拉格朗日求解器可实现对任意模型或计算条件的无条件稳定。构建了BP神经网络用于预测水滴撞击概率,通过筛除非撞击水滴减少计算量。BP神经网络不针对特定模型提前训练,使用异步并行策略使BP神经网络训练和水滴运动同时求解,建立了广泛适用的异步拉格朗日求解器。使用GLC-305后掠三维翼型和某型发动机短舱模型对求解器进行验证,结果显示BP神经网络可以有效提升计算效率,对比没有神经网络辅助最多节省27%运行时间,同时保有同等计算精确度。本文研究为首次尝试神经网络技术与结冰数值模拟融合,为进一步发展拉格朗日方法提供有力支撑。In-flight icing is threatening aviation safety.The Lagrangian method is widely used in aircraft icing simulation to solve water collection efficiency,the development of which has been impeded by robustness issues and high computational cost.To resolve these disadvantages,two critical algorithms are employed in this study.The Monte Carlo integral method is applied to calculate collection efficiency,which makes the Lagrangian method unconditionally robust for an arbitrary situation.The backpropagation(BP)neural network is also implanted to make a rapid prediction of droplet impingement.Additionally,these two algorithms are deeply coupled in an asynchronous parallelism that allows un-interfered parallel for each procedure respectively.The current study is implemented in NNW-ICE software platform.The asynchronous solver is evaluated with a 3D GLC-305 airfoil and a jet engine nacelle model.The result shows that the BP network contributes a significant acceleration to the Monte Carlo method,saving about 27%running time to achieve equal accurate result.The study is a first attempt for coupling the neural network art and numerical simulation in aircraft icing,providing strong support for the improvement of Lagrangian method and aircraft icing.

关 键 词：飞机结冰 拉格朗日方法 蒙特卡洛方法 BP神经网络 异步并行 

分 类 号：O359.1[理学—流体力学]