[{"data":1,"prerenderedAt":156},["ShallowReactive",2],{"content-query-r7qkVUfG6P":3},{"_path":4,"_dir":5,"_draft":6,"_partial":6,"_locale":7,"title":8,"description":9,"date":10,"cover":11,"type":12,"body":13,"_type":150,"_id":151,"_source":152,"_file":153,"_stem":154,"_extension":155},"/news/en/2759","en",false,"","PeRCNN: A New AI + Scientific Computing Achievement Published in Nature","Recently, Huawei, collaborating with Professor Sun Hao's team from the Gaoling School of Artificial Intelligence, Renmin University of China, proposed a new recurrent network called PeRCNN.","2023-07-24","https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2023/09/14/1e159140aa5e4682a4322d4577eacd11.png","news",{"type":14,"children":15,"toc":147},"root",[16,24,38,43,48,56,61,66,71,78,83,88,95,100,105,110,119,132,137,142],{"type":17,"tag":18,"props":19,"children":21},"element","h1",{"id":20},"percnn-a-new-ai-scientific-computing-achievement-published-in-nature",[22],{"type":23,"value":8},"text",{"type":17,"tag":25,"props":26,"children":27},"p",{},[28,30,36],{"type":23,"value":29},"Recently, Huawei, collaborating with Professor Sun Hao's team from the Gaoling School of Artificial Intelligence, Renmin University of China, proposed a new recurrent network called PeRCNN. This physics-encoded recurrent convolutional neural network is developed based on the Ascend basic software and hardware platform and MindSpore framework. This achievement has been published in ",{"type":17,"tag":31,"props":32,"children":33},"em",{},[34],{"type":23,"value":35},"Nature Machine Intelligence",{"type":23,"value":37},", and its code has been open-sourced in the MindFlow code repository on Gitee[1].",{"type":17,"tag":25,"props":39,"children":40},{},[41],{"type":23,"value":42},"Compared with methods such as physics-informed neural networks (PINNs), ConvLSTM, and PDE-NET, PeRCNN outperforms in model generalization and anti-noise and improves the long-term inference precision by more than 10 times, having great potential in aerospace, ship manufacturing, meteorological forecast, and other fields.",{"type":17,"tag":25,"props":44,"children":45},{},[46],{"type":23,"value":47},"The partial differential equation (PDE) plays a central role in the modeling of physical systems. However, in fields such as epidemiology, meteorological science, fluid mechanics, and biology, many underlying PDEs have not been fully discovered. Additionally, it requires a large amount of computing power to accurately calculate existing PDEs, such as the Navier-Stokes equations, which poses challenges for numerical simulation in large-scale spatiotemporal systems. At present, the advancement of machine learning provides new solutions and inversion methods for PDEs.",{"type":17,"tag":25,"props":49,"children":50},{},[51],{"type":17,"tag":52,"props":53,"children":55},"img",{"alt":7,"src":54},"https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2023/09/14/2ed0e330cc854450b2397d04b84c2c49.png",[],{"type":17,"tag":25,"props":57,"children":58},{},[59],{"type":23,"value":60},"PeRCNN model framework",{"type":17,"tag":25,"props":62,"children":63},{},[64],{"type":23,"value":65},"Current data-driven models rely on big data[2], which is difficult to obtain for most scientific issues and incurs explainability problems. Although PINNs[3] use prior knowledge to constrain model training and reduce dependency on data, their soft constraints are determined by loss functions, limiting the accuracy of final results. As a result, the pursuit of achieving results with high accuracy, robustness, explainability, and generalization in the absence of valid data remains a challenge for the academic community.",{"type":17,"tag":25,"props":67,"children":68},{},[69],{"type":23,"value":70},"With this context, Huawei and Professor Sun Hao's team developed Ascend- and MindSpore-based PeRCNN[4] that can accurately obtain approximate results of nonlinear PDEs.",{"type":17,"tag":25,"props":72,"children":73},{},[74],{"type":17,"tag":52,"props":75,"children":77},{"alt":7,"src":76},"https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2023/09/14/d78b08e0f47c49f7a38895571b189f7f.png",[],{"type":17,"tag":25,"props":79,"children":80},{},[81],{"type":23,"value":82},"In terms of applications of reaction-diffusion equations, PeRCNN has a better effect than ConvLSTM and PINN.",{"type":17,"tag":25,"props":84,"children":85},{},[86],{"type":23,"value":87},"The PeRCNN neural network forcibly encodes the physical structure. Based on the π-convolution module designed with part of physical prior knowledge, PeRCNN implements nonlinear approximation by multiplying elements among feature maps. The physical encoding mechanism ensures that the model strictly obeys the given physical equation according to the prior knowledge. This proposed method is applied to various issues related to PDE systems, including data-driven modeling and PDE discovery, and ensures accuracy and generalization.",{"type":17,"tag":25,"props":89,"children":90},{},[91],{"type":17,"tag":52,"props":92,"children":94},{"alt":7,"src":93},"https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2023/09/14/3b0bdab65f8a42cc93b8339ede14e962.png",[],{"type":17,"tag":25,"props":96,"children":97},{},[98],{"type":23,"value":99},"Also, PeRCNN outperforms ConvLSTM, ResNET, PDE-NET, and DHPM in prediction and extrapolation.",{"type":17,"tag":25,"props":101,"children":102},{},[103],{"type":23,"value":104},"Another unique advantage of PeRCNN is its explainability due to the multiplication form of π-convolution. Symbolic calculation is able to further extract basic physics expressions at bottom layers from the learned model. This allows PeRCNN to be used as an effective tool to help developers accurately and reliably discover potential physical laws from imperfect and highly noisy data.",{"type":17,"tag":25,"props":106,"children":107},{},[108],{"type":23,"value":109},"Strong nonlinear phenomena such as turbulence and shock waves often occur in disciplines such as fluid mechanics, meteorology, and oceanography. Traditional numerical methods require a large number of computing resources. Currently, AI has shown great potential for problems such as aircraft flow fields and mid-term weather forecast. PeRCNN, with its strong generalization, anti-noise, and high accuracy, is poised to break computing bottlenecks in various fields. It is expected to accelerate industrial simulation and design and become a powerful new tool in the AI + scientific computing field.",{"type":17,"tag":25,"props":111,"children":112},{},[113],{"type":17,"tag":114,"props":115,"children":116},"strong",{},[117],{"type":23,"value":118},"References",{"type":17,"tag":25,"props":120,"children":121},{},[122,124],{"type":23,"value":123},"[1]",{"type":17,"tag":125,"props":126,"children":130},"a",{"href":127,"rel":128},"https://gitee.com/mindspore/mindscience/tree/master/MindFlow/applications/data_mechanism_fusion/PeRCNN",[129],"nofollow",[131],{"type":23,"value":127},{"type":17,"tag":25,"props":133,"children":134},{},[135],{"type":23,"value":136},"[2]Yann LeCun, Yoshua Bengio, and Geoffrey Hinton. Deep learning. Nature, 521(7553):436–444, 2015.",{"type":17,"tag":25,"props":138,"children":139},{},[140],{"type":23,"value":141},"[3]Maziar Raissi, Paris Perdikaris, and George E Karniadakis. Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations. Journal of Computational Physics, 378:686–707, 2019.",{"type":17,"tag":25,"props":143,"children":144},{},[145],{"type":23,"value":146},"[4]Chengping Rao, Pu Ren, Qi Wang, Oral Buyukozturk, Hao Sun*, Yang Liu*. Encoding physics to learn reaction-diffusion processes. Nature Machine Intelligence, 2023, DOI: 10.1038/s42256-023-00685-7",{"title":7,"searchDepth":148,"depth":148,"links":149},4,[],"markdown","content:news:en:2759.md","content","news/en/2759.md","news/en/2759","md",1776506045713]