[{"data":1,"prerenderedAt":313},["ShallowReactive",2],{"content-query-tk8NYFjtS6":3},{"_path":4,"_dir":5,"_draft":6,"_partial":6,"_locale":7,"title":8,"description":9,"date":10,"cover":11,"type":12,"category":13,"body":14,"_type":307,"_id":308,"_source":309,"_file":310,"_stem":311,"_extension":312},"/technology-blogs/zh/1926","zh",false,"","论文精讲 | 神经网络从数据中学习空气动力学物理定律","该工作表明神经网络可以从数据中学习并提取物理定律，用于物理任务的神经网络有可解释性。","2022-10-25","https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2022/11/01/23099a6eaf674c57ba0929f0b914e364.png","technology-blogs","大V博文",{"type":15,"children":16,"toc":302},"root",[17,25,34,39,44,73,80,85,90,97,102,109,114,125,132,137,144,149,156,161,166,173,178,182,192,197,212,220,227,235,245,261,271,287],{"type":18,"tag":19,"props":20,"children":22},"element","h1",{"id":21},"论文精讲-神经网络从数据中学习空气动力学物理定律",[23],{"type":24,"value":8},"text",{"type":18,"tag":26,"props":27,"children":28},"p",{},[29],{"type":18,"tag":30,"props":31,"children":33},"img",{"alt":7,"src":32},"https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2022/11/01/06702a86176e4b139a3affb7746330f6.gif",[],{"type":18,"tag":26,"props":35,"children":36},{},[37],{"type":24,"value":38},"机器学习和基础学科交叉在近年受到越来越多的关注。能够从大量数据中学习的 AI，是否能够像人类一样，从数据中发现规律？当神经网络被用于解决物理问题时，是否有可能学习到物理知识？",{"type":18,"tag":26,"props":40,"children":41},{},[42],{"type":24,"value":43},"近日，来自加拿大蒙特利尔大学的研究人员给出了肯定的回答。该工作在研究 AI 如何完成翼型气动系数预测任务过程中，发现了神经网络在训练过程中学习空气动力学物理定律的证据。",{"type":18,"tag":26,"props":45,"children":46},{},[47,49,55,57,62,64,71],{"type":24,"value":48},"该研究以《",{"type":18,"tag":50,"props":51,"children":52},"em",{},[53],{"type":24,"value":54},"Learning aerodynamics with neural network",{"type":24,"value":56},"》为题，于 2022 年 4月发布在《",{"type":18,"tag":50,"props":58,"children":59},{},[60],{"type":24,"value":61},"Scientific Reports",{"type":24,"value":63},"》上。该文章的第一兼通讯作者彭文辉通讯地址为",{"type":18,"tag":65,"props":66,"children":68},"a",{"href":67},"mailto:_wenhui.peng@polymtl.ca",[69],{"type":24,"value":70},"_wenhui.peng@polymtl.ca",{"type":24,"value":72},"._",{"type":18,"tag":26,"props":74,"children":75},{},[76],{"type":18,"tag":30,"props":77,"children":79},{"alt":7,"src":78},"https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2022/11/01/904f2327f24647019ae86bccfba2d221.png",[],{"type":18,"tag":26,"props":81,"children":82},{},[83],{"type":24,"value":84},"计算翼型气动系数是空气动力学中的关键任务之一，该任务通常由计算流体力学方法完成。近年来，机器学习的快速发展为计算流体力学带来了新的机遇。为了进一步提高计算效率和预测精度，",{"type":18,"tag":26,"props":86,"children":87},{},[88],{"type":24,"value":89},"研究人员提出了一种神经网络模型——空间单元卷积神经网络（Element Spatial Convolution Neural Network ，ESCNN），用于预测翼型升力系数。ESCNN 的输入为翼型坐标和攻角，输出为升力系数，其架构如图 1 所示。",{"type":18,"tag":26,"props":91,"children":92},{},[93],{"type":18,"tag":30,"props":94,"children":96},{"alt":7,"src":95},"https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2022/11/01/1b3ff43f4c8448979badb32017f01385.png",[],{"type":18,"tag":26,"props":98,"children":99},{},[100],{"type":24,"value":101},"相较于现有 SOTA 模型，ESCNN 减少了两个数量级的参数，但取得了更高的预测精度。为了调查该神经网络如何实现高精确预测，研究人员对训练中的 ESCNN 进行调查：他们追踪了图 1 中的全连接层，并记录了其中神经元 [f1,......fn] 在训练过程中的进化过程，如图 2 所示。",{"type":18,"tag":26,"props":103,"children":104},{},[105],{"type":18,"tag":30,"props":106,"children":108},{"alt":7,"src":107},"https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2022/11/01/066a48e5876843c8b13c2c7d54ab81b6.png",[],{"type":18,"tag":26,"props":110,"children":111},{},[112],{"type":24,"value":113},"他们发现了两个有趣的现象：（1）随着训练过程进行，该全连接层最终收敛为一个正弦波形曲线；（2）在进化过程中，该全连接层的第一个神经元 f1 和最后一个神经元 fn 始终保持数值大小一致。",{"type":18,"tag":26,"props":115,"children":116},{},[117,119,123],{"type":24,"value":118},"库达条件（Kutta condition）是空气动力学中的基本物理定律之一：翼型尾部的上下涡流强度必须相等，以保证气体平滑流过翼型，即",{"type":18,"tag":30,"props":120,"children":122},{"alt":7,"src":121},"https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2022/11/01/1d4845f7f478465ab6086604c014c5ca.png",[],{"type":24,"value":124},"，如图 3 所示。全连接层在进化过程中这种 f1=fn 的自我约束是神经网络学习到库达条件的证据。",{"type":18,"tag":26,"props":126,"children":127},{},[128],{"type":18,"tag":30,"props":129,"children":131},{"alt":7,"src":130},"https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2022/11/01/ccc2c95cbc7249e28c758945afb680b3.png",[],{"type":18,"tag":26,"props":133,"children":134},{},[135],{"type":24,"value":136},"另一方面，随着训练进行，为什么全连接层会收敛成为类正弦曲线？研究人员计算并对比了翼型的涡强分布和全连接层的神经元数值，发现他们是高度相关的，如图4 所示。这是神经网络学习到物理量涡强的证据。",{"type":18,"tag":26,"props":138,"children":139},{},[140],{"type":18,"tag":30,"props":141,"children":143},{"alt":7,"src":142},"https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2022/11/01/d595c2f42c03405e931e8334e8336aa9.png",[],{"type":18,"tag":26,"props":145,"children":146},{},[147],{"type":24,"value":148},"涡强有正有负，如果通过 ReLU 激活函数来限制全连接层神经元只输出非负值（图5 所示），神经网络能否依然学到涡强分布？",{"type":18,"tag":26,"props":150,"children":151},{},[152],{"type":18,"tag":30,"props":153,"children":155},{"alt":7,"src":154},"https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2022/11/01/3cc90617c27d4e1eb1e5a4bdac8b1039.png",[],{"type":18,"tag":26,"props":157,"children":158},{},[159],{"type":24,"value":160},"研究人员发现了另一现象：虽然 ReLU 只限制了神经元输出非负数（图 5c），神经网络在进化过程中会自动限制正数域的输出区间，以获得更好的对称性（图5def）。为什么神经网络偏好对称性，对称性有什么意义？",{"type":18,"tag":26,"props":162,"children":163},{},[164],{"type":24,"value":165},"研究人员从训练翼型数据中提取了一些随机样本：不同翼型形状不同，但翼型坐标的间隔分布基本对称（图 6 所示）。这是神经网络学习到翼型几何特性的证据。",{"type":18,"tag":26,"props":167,"children":168},{},[169],{"type":18,"tag":30,"props":170,"children":172},{"alt":7,"src":171},"https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2022/11/01/1bcb948db59d429cb76f448c8a9af8a6.png",[],{"type":18,"tag":26,"props":174,"children":175},{},[176],{"type":24,"value":177},"该工作提出了一种用于预测翼型升力系数的神经网络，相较于现有 SOTA 模型减少两个数量级的参数，但实现了更高预测精度。在研究神经网络如何进行推理的过程中，研究人员发现了神经网络自学库达条件，翼型涡强分布和几何对称性的证据。",{"type":18,"tag":26,"props":179,"children":180},{},[181],{"type":24,"value":9},{"type":18,"tag":183,"props":184,"children":186},"h2",{"id":185},"致谢",[187],{"type":18,"tag":188,"props":189,"children":190},"strong",{},[191],{"type":24,"value":185},{"type":18,"tag":26,"props":193,"children":194},{},[195],{"type":24,"value":196},"本研究成果得到了中国人工智能学会-华为MindSpore学术奖励基金的资助。后面会出一期该论文的MindSpore代码解读，敬请期待，谢谢！",{"type":18,"tag":26,"props":198,"children":199},{},[200],{"type":18,"tag":50,"props":201,"children":202},{},[203,205],{"type":24,"value":204},"论文地址：",{"type":18,"tag":65,"props":206,"children":210},{"href":207,"rel":208},"https://www.nature.com/articles/s41598-022-10737-4",[209],"nofollow",[211],{"type":24,"value":207},{"type":18,"tag":26,"props":213,"children":214},{},[215],{"type":18,"tag":50,"props":216,"children":217},{},[218],{"type":24,"value":219},"作者信息：彭文辉(Wenhui Peng), Yao Zhang, Eric Laurendeau, Michel Desmarais",{"type":18,"tag":26,"props":221,"children":222},{},[223],{"type":18,"tag":30,"props":224,"children":226},{"alt":7,"src":225},"https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2022/10/24/f5482b1c286f4581b7f3deb4b69c8d1a.jpg",[],{"type":18,"tag":26,"props":228,"children":229},{},[230],{"type":18,"tag":188,"props":231,"children":232},{},[233],{"type":24,"value":234},"MindSpore官方资料",{"type":18,"tag":26,"props":236,"children":237},{},[238,243],{"type":18,"tag":188,"props":239,"children":240},{},[241],{"type":24,"value":242},"官方QQ群",{"type":24,"value":244}," : 486831414",{"type":18,"tag":26,"props":246,"children":247},{},[248,253,255],{"type":18,"tag":188,"props":249,"children":250},{},[251],{"type":24,"value":252},"官网",{"type":24,"value":254},"：",{"type":18,"tag":65,"props":256,"children":259},{"href":257,"rel":258},"https://www.mindspore.cn/",[209],[260],{"type":24,"value":257},{"type":18,"tag":26,"props":262,"children":263},{},[264,269],{"type":18,"tag":188,"props":265,"children":266},{},[267],{"type":24,"value":268},"Gitee",{"type":24,"value":270}," : https : //gitee.com/mindspore/mindspore",{"type":18,"tag":26,"props":272,"children":273},{},[274,279,281],{"type":18,"tag":188,"props":275,"children":276},{},[277],{"type":24,"value":278},"GitHub",{"type":24,"value":280}," : ",{"type":18,"tag":65,"props":282,"children":285},{"href":283,"rel":284},"https://github.com/mindspore-ai/mindspore",[209],[286],{"type":24,"value":283},{"type":18,"tag":26,"props":288,"children":289},{},[290,295,296],{"type":18,"tag":188,"props":291,"children":292},{},[293],{"type":24,"value":294},"论坛",{"type":24,"value":254},{"type":18,"tag":65,"props":297,"children":300},{"href":298,"rel":299},"https://bbs.huaweicloud.com/forum/forum-1076-1.html",[209],[301],{"type":24,"value":298},{"title":7,"searchDepth":303,"depth":303,"links":304},4,[305],{"id":185,"depth":306,"text":185},2,"markdown","content:technology-blogs:zh:1926.md","content","technology-blogs/zh/1926.md","technology-blogs/zh/1926","md",1776506117060]