[{"data":1,"prerenderedAt":492},["ShallowReactive",2],{"content-query-m4MTN7ntcY":3},{"_path":4,"_dir":5,"_draft":6,"_partial":6,"_locale":7,"title":8,"description":9,"date":10,"cover":11,"type":12,"body":13,"_type":486,"_id":487,"_source":488,"_file":489,"_stem":490,"_extension":491},"/technology-blogs/zh/3808","zh",false,"","基于昇思MindSpore，北京航空航天大学科研团队提出科学智算基础模型OmniArch，实现11类PDE求解性能提升","“大自然的书籍是用数学语言写成的。”——伽利略","2025-08-04","https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2025/08/08/9dfa44d254fe44008828ac6213056060.png","technology-blogs",{"type":14,"children":15,"toc":483},"root",[16,24,29,49,57,62,70,83,94,108,123,130,135,140,157,162,176,181,189,194,201,206,211,216,221,229,234,241,256,268,276,283,288,306,311,325,330,343,350,358,363,368,380,393,400,424,431,448,461,468,473,478],{"type":17,"tag":18,"props":19,"children":21},"element","h1",{"id":20},"基于昇思mindspore北京航空航天大学科研团队提出科学智算基础模型omniarch实现11类pde求解性能提升",[22],{"type":23,"value":8},"text",{"type":17,"tag":25,"props":26,"children":27},"p",{},[28],{"type":23,"value":9},{"type":17,"tag":25,"props":30,"children":31},{},[32,34,40,42,47],{"type":23,"value":33},"而",{"type":17,"tag":35,"props":36,"children":37},"strong",{},[38],{"type":23,"value":39},"偏微分方程（PDE）",{"type":23,"value":41},"，正是这本著作中最核心的篇章之一。它刻画了从量子涨落到星系演化，从流体冲击到电磁传播的",{"type":17,"tag":35,"props":43,"children":44},{},[45],{"type":23,"value":46},"普适物理规律",{"type":23,"value":48},"。然而，求解PDE却如同破译大自然的密码，传统数值方法计算昂贵，专用神经网络模型则局限特定问题，一个模型通常只能求解一种维度、一类物理量、特定边界条件的方程。面对真实世界多尺度、多物理场强耦合的复杂系统，人们不得不为每个子问题“量身定制”求解器，知识无法迁移与复用，模型也难以通过数据和参数进行规模扩展。",{"type":17,"tag":25,"props":50,"children":51},{},[52],{"type":17,"tag":35,"props":53,"children":54},{},[55],{"type":23,"value":56},"能否像ChatGPT理解语言一样，构建一个理解并求解“物理语言”的通用基础模型？",{"type":17,"tag":25,"props":58,"children":59},{},[60],{"type":23,"value":61},"近期，北京航空航天大学李建欣教授团队提出了科学智算基础模型OmniArch，首次实现单个模型统一求解1D、2D、3D PDE的重大突破，并在多个关键指标上大幅超越现有方法。相关成果以《OmniArch: Building Foundation Model For Scientific Computing》为题，被ICML 2025接收。",{"type":17,"tag":25,"props":63,"children":64},{},[65],{"type":17,"tag":66,"props":67,"children":69},"img",{"alt":7,"src":68},"https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2025/08/08/84435562589a41e28973a82d9d3458b6.png",[],{"type":17,"tag":25,"props":71,"children":72},{},[73,75],{"type":23,"value":74},"论文地址：",{"type":17,"tag":76,"props":77,"children":81},"a",{"href":78,"rel":79},"https://openreview.net/pdf?id=UlprLwWYKP",[80],"nofollow",[82],{"type":23,"value":78},{"type":17,"tag":25,"props":84,"children":85},{},[86,88],{"type":23,"value":87},"代码地址：",{"type":17,"tag":76,"props":89,"children":92},{"href":90,"rel":91},"https://openi.pcl.ac.cn/cty315/OmniArch",[80],[93],{"type":23,"value":90},{"type":17,"tag":25,"props":95,"children":96},{},[97,102],{"type":17,"tag":35,"props":98,"children":99},{},[100],{"type":23,"value":101},"CNAI4S科学智算共性平台：",{"type":17,"tag":76,"props":103,"children":106},{"href":104,"rel":105},"https://cnai4s.com/",[80],[107],{"type":23,"value":104},{"type":17,"tag":25,"props":109,"children":110},{},[111,116,118],{"type":17,"tag":35,"props":112,"children":113},{},[114],{"type":23,"value":115},"# 01",{"type":23,"value":117}," ",{"type":17,"tag":35,"props":119,"children":120},{},[121],{"type":23,"value":122},"挑战：阻碍“物理GPT”-OmniArch的三座大山",{"type":17,"tag":25,"props":124,"children":125},{},[126],{"type":17,"tag":66,"props":127,"children":129},{"alt":7,"src":128},"https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2025/08/08/2e14ce5370e3419cb19ce2e84e19dae3.png",[],{"type":17,"tag":25,"props":131,"children":132},{},[133],{"type":23,"value":134},"构建面向PDE求解的通用基础模型，需跨越三大核心障碍：",{"type":17,"tag":25,"props":136,"children":137},{},[138],{"type":23,"value":139},"**1. 多维度统一表示：**物理数据形态各异（1D序列、2D场、3D体），传统架构需为不同维度设计专门网络，知识难以共享。",{"type":17,"tag":25,"props":141,"children":142},{},[143,155],{"type":17,"tag":35,"props":144,"children":145},{},[146,148,153],{"type":23,"value":147},"2. 多物理量统一处理：",{"type":17,"tag":35,"props":149,"children":150},{},[151],{"type":23,"value":152},"不同PDE包含的物理量数量迥异（如流体方程含速度/压力/密度，电磁方程有6个分量）。统一模型需能",{"type":23,"value":154},"动态处理",{"type":23,"value":156},"变化的物理量组合及其复杂相互作用。",{"type":17,"tag":25,"props":158,"children":159},{},[160],{"type":23,"value":161},"**3. 物理约束对齐：**每个PDE都蕴含各自的“物理法则”（守恒律、边界条件、对称性）。基础模型如何在保持灵活性的同时，自适应地满足这些严格约束，确保解的物理合理性？",{"type":17,"tag":25,"props":163,"children":164},{},[165,170,171],{"type":17,"tag":35,"props":166,"children":167},{},[168],{"type":23,"value":169},"# 02",{"type":23,"value":117},{"type":17,"tag":35,"props":172,"children":173},{},[174],{"type":23,"value":175},"OmniArch：统一架构的三大创新突破",{"type":17,"tag":25,"props":177,"children":178},{},[179],{"type":23,"value":180},"OmniArch创造性地融合了频域变换、注意力机制与对比学习，攻克了上述难题：",{"type":17,"tag":25,"props":182,"children":183},{},[184],{"type":17,"tag":35,"props":185,"children":186},{},[187],{"type":23,"value":188},"1、Fourier编码器实现维度统一",{"type":17,"tag":25,"props":190,"children":191},{},[192],{"type":23,"value":193},"尽管物理数据在欧氏空间维度不同，但在频域空间具有统一的多尺度结构。因此，OmniArch对输入坐标和物理场进行快速傅里叶变换（FFT），保留最重要的Top-K频率分量（模式），从而将不同维度/分辨率的数据转换为统一长度的频域表示。",{"type":17,"tag":25,"props":195,"children":196},{},[197],{"type":17,"tag":66,"props":198,"children":200},{"alt":7,"src":199},"https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2025/08/08/7d93459b46874ac39d5c512787ac46cd.png",[],{"type":17,"tag":25,"props":202,"children":203},{},[204],{"type":23,"value":205},"优势：",{"type":17,"tag":25,"props":207,"children":208},{},[209],{"type":23,"value":210},"(1) 计算效率从O(N²)降至O(NlogN)；",{"type":17,"tag":25,"props":212,"children":213},{},[214],{"type":23,"value":215},"(2) 低频捕获全局模式，高频编码局部细节；",{"type":17,"tag":25,"props":217,"children":218},{},[219],{"type":23,"value":220},"(3) 不同分辨率数据在低频部分自然对齐。",{"type":17,"tag":25,"props":222,"children":223},{},[224],{"type":17,"tag":35,"props":225,"children":226},{},[227],{"type":23,"value":228},"2、Temporal Mask实现多物理变量耦合",{"type":17,"tag":25,"props":230,"children":231},{},[232],{"type":23,"value":233},"标准Transformer采用因果注意力机制，但这会破坏同一时刻物理量之间的相互作用。OmniArch引入Temporal Mask：",{"type":17,"tag":25,"props":235,"children":236},{},[237],{"type":17,"tag":66,"props":238,"children":240},{"alt":7,"src":239},"https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2025/08/08/129b74abbb374314878d57de41128bcb.png",[],{"type":17,"tag":242,"props":243,"children":244},"ul",{},[245,251],{"type":17,"tag":246,"props":247,"children":248},"li",{},[249],{"type":23,"value":250},"时刻内完全连接：同一时间步的所有物理量token可相互“看见”并交互，精确建模耦合关系（如NS方程中的速度-压力耦合）。",{"type":17,"tag":246,"props":252,"children":253},{},[254],{"type":23,"value":255},"时刻间因果约束： 严格遵循时间序列因果性，未来状态不影响过去。",{"type":17,"tag":25,"props":257,"children":258},{},[259,261,266],{"type":23,"value":260},"这种设计自然地编码了物理系统",{"type":17,"tag":35,"props":262,"children":263},{},[264],{"type":23,"value":265},"时序演进与瞬时耦合",{"type":23,"value":267},"的核心特性。",{"type":17,"tag":25,"props":269,"children":270},{},[271],{"type":17,"tag":35,"props":272,"children":273},{},[274],{"type":23,"value":275},"3、PDE-Aligner增强物理一致性",{"type":17,"tag":25,"props":277,"children":278},{},[279],{"type":17,"tag":66,"props":280,"children":282},{"alt":7,"src":281},"https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2025/08/08/cfd67a22ae124d019838e18755ca37b6.png",[],{"type":17,"tag":25,"props":284,"children":285},{},[286],{"type":23,"value":287},"在微调阶段，PDE-Aligner通过对比学习将符号方程与数值演化对齐：",{"type":17,"tag":242,"props":289,"children":290},{},[291,296,301],{"type":17,"tag":246,"props":292,"children":293},{},[294],{"type":23,"value":295},"方程编码：预训练BERT处理LaTeX格式的PDE文本",{"type":17,"tag":246,"props":297,"children":298},{},[299],{"type":23,"value":300},"物理特征：频域相位差Δφ和振幅比R编码演化模式",{"type":17,"tag":246,"props":302,"children":303},{},[304],{"type":23,"value":305},"能量约束：确保∑|R|≈1满足Parseval定理",{"type":17,"tag":25,"props":307,"children":308},{},[309],{"type":23,"value":310},"这种“软对齐”避免了硬编码物理约束，使模型能够自适应不同方程和边界条件的约束。",{"type":17,"tag":25,"props":312,"children":313},{},[314,319,320],{"type":17,"tag":35,"props":315,"children":316},{},[317],{"type":23,"value":318},"# 03",{"type":23,"value":117},{"type":17,"tag":35,"props":321,"children":322},{},[323],{"type":23,"value":324},"实验验证",{"type":17,"tag":25,"props":326,"children":327},{},[328],{"type":23,"value":329},"OmniArch在涵盖流体力学、电磁学、反应扩散等领域的11种经典PDE上进行了严格测试（数据集：PDEBench, PDEArena），对比了包括专用模型（PINNs, U-Net, FNO）和前沿统一预训练模型（MPP, Poseidon, DPOT, PDEformer-1）在内的众多基线。",{"type":17,"tag":25,"props":331,"children":332},{},[333,335],{"type":23,"value":334},"****1、****",{"type":17,"tag":35,"props":336,"children":337},{},[338],{"type":17,"tag":35,"props":339,"children":340},{},[341],{"type":23,"value":342},"基准性能：精度飞跃，最高提升98.7%",{"type":17,"tag":25,"props":344,"children":345},{},[346],{"type":17,"tag":66,"props":347,"children":349},{"alt":7,"src":348},"https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2025/08/08/9baf059142ed41a096a7c90fd69ec367.png",[],{"type":17,"tag":25,"props":351,"children":352},{},[353],{"type":17,"tag":35,"props":354,"children":355},{},[356],{"type":23,"value":357},"表：OmniArch在11类PDE上的nRMSE性能对比（越低越好），显著优于专用模型（上）与其他统一预训练模型（中）。下划线为之前SOTA，加粗为本文最佳。",{"type":17,"tag":25,"props":359,"children":360},{},[361],{"type":23,"value":362},"**One vs All：**OmniArch（Base/Large版本）**在全****部11项测试中均取得最佳或接近最佳性能。**单个模型、单份权重即可全面替代针对不同维度、不同PDE专门设计的众多模型。",{"type":17,"tag":25,"props":364,"children":365},{},[366],{"type":23,"value":367},"**CFD求解的显著优势：**在最具挑战性的计算流体动力学（CFD）任务上，OmniArch展现出较大优势，其中1D CFD方程求解相比最佳基线，误差降低 98.7%。",{"type":17,"tag":25,"props":369,"children":370},{},[371,373,378],{"type":23,"value":372},"**物理对齐效果：",{"type":17,"tag":35,"props":374,"children":375},{},[376],{"type":23,"value":377},"PDE-Aligner微调带来",{"type":23,"value":379},"平均超过20%**的精度提升，证明了其有效注入物理知识的能力。",{"type":17,"tag":25,"props":381,"children":382},{},[383,385],{"type":23,"value":384},"****2、****",{"type":17,"tag":35,"props":386,"children":387},{},[388],{"type":17,"tag":35,"props":389,"children":390},{},[391],{"type":23,"value":392},"零样本泛化：“举一反三”的物理直觉",{"type":17,"tag":25,"props":394,"children":395},{},[396],{"type":17,"tag":66,"props":397,"children":399},{"alt":7,"src":398},"https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2025/08/08/1e66fc8ee5a9465284b384f20d8cdd05.png",[],{"type":17,"tag":25,"props":401,"children":402},{},[403,415,417,422],{"type":17,"tag":35,"props":404,"children":405},{},[406,408,413],{"type":23,"value":407},"未见过的方程直接解：",{"type":17,"tag":35,"props":409,"children":410},{},[411],{"type":23,"value":412},"OmniArch无需微调，即可对",{"type":23,"value":414},"训练集之外",{"type":23,"value":416},"的复杂",{"type":17,"tag":35,"props":418,"children":419},{},[420],{"type":23,"value":421},"PDE",{"type":23,"value":423},"进行高精度求解。在Shock（激波）、KH（开尔文-亥姆霍兹不稳定性）、OTVortex等三个高难度2D流体物理系统上，其零样本性能显著超越了专门在这些任务上训练的FNO、U-Net模型。相比同样做预训练的MPP-Large模型，误差降低了4-7倍 (如KH: 0.2763 vs. 1.3261)。",{"type":17,"tag":25,"props":425,"children":426},{},[427],{"type":17,"tag":66,"props":428,"children":430},{"alt":7,"src":429},"https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2025/08/08/e04a47497674484bb7fa6628e736ae08.png",[],{"type":17,"tag":25,"props":432,"children":433},{},[434,446],{"type":17,"tag":35,"props":435,"children":436},{},[437,439,444],{"type":23,"value":438},"空间分辨率直接泛化：",{"type":17,"tag":35,"props":440,"children":441},{},[442],{"type":23,"value":443},"得益于频域表示的尺度不变性，OmniArch在训练（如128x128）后，无需修改架构即可",{"type":23,"value":445},"直接处理",{"type":23,"value":447},"截然不同的输入分辨率（32x32, 256x256, 512x512），且精度保持在可接受范围（nRMSE \u003C 0.2），计算延迟几乎不变。这解决了传统方法因网格绑定需重新训练或插值的问题。",{"type":17,"tag":25,"props":449,"children":450},{},[451,453],{"type":23,"value":452},"****3、****",{"type":17,"tag":35,"props":454,"children":455},{},[456],{"type":17,"tag":35,"props":457,"children":458},{},[459],{"type":23,"value":460},"上下文学习：动态“领悟”物理规则",{"type":17,"tag":25,"props":462,"children":463},{},[464],{"type":17,"tag":66,"props":465,"children":467},{"alt":7,"src":466},"https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2025/08/08/9261b0fbd6034403bc4b7dd88494378d.png",[],{"type":17,"tag":25,"props":469,"children":470},{},[471],{"type":23,"value":472},"如同大语言模型（LLM）通过提示（Prompt）理解新任务，经过预训练的OmniArch展现出了上下文学习能力。给定少量历史时间步的观测数据（称为“PDE Prompt”），模型即可动态学习并模拟出该特定PDE的算子。",{"type":17,"tag":25,"props":474,"children":475},{},[476],{"type":23,"value":477},"依托昇思MindSpore AI框架的开发训练微调部署全方位支持，以及昇腾AI处理器强大的计算能力，OmniArch在混合精度训练中实现了显著的性能提升。对于LayerNorm、SoftMax等关键算子，MindSpore支持自动提升至FP32精度，其他模型参数和梯度计算采用FP16半精度存储，在保证数值稳定性的同时，训练速度提升约40%，显存占用减少近50%。OmniArch基于昇腾软硬件高度协同优化的傅里叶变换等算子，已加速赋能MindSpore Flow等新一代流体力学科学计算套件，并有望实现自主创新软硬件生态能力的最大化释放。",{"type":17,"tag":25,"props":479,"children":480},{},[481],{"type":23,"value":482},"未来，团队计划将OmniArch拓展到更广泛的PDE类型、更复杂的边界条件，并探索其在气候模拟、航天航空器辅助设计等重大科学工程问题中的落地应用。",{"title":7,"searchDepth":484,"depth":484,"links":485},4,[],"markdown","content:technology-blogs:zh:3808.md","content","technology-blogs/zh/3808.md","technology-blogs/zh/3808","md",1776506135599]