[{"data":1,"prerenderedAt":275},["ShallowReactive",2],{"content-query-9tC68gUXCA":3},{"_path":4,"_dir":5,"_draft":6,"_partial":6,"_locale":7,"title":8,"description":9,"date":10,"cover":11,"type":12,"body":13,"_type":269,"_id":270,"_source":271,"_file":272,"_stem":273,"_extension":274},"/news/zh/3070","zh",false,"","任意PDE的秒级求解？昇思MindSpore最新成果PDEformer-1迎来开源！","近日，华为AI4SCI Lab联合北京大学北京国际数学研究中心教授、北京大学国际机器学习研究中心副主任董彬教授团队，在昇腾AI处理器的强大算力支持下，基于全场景AI框架昇思MindSpore推出了一维含时偏微分方程通用模型PDEformer-1。该成果在昇思人工智能框架峰会2024上首次发布亮相。","2024-04-15","https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2024/04/19/0ec7b46611064702921564728b7ec854.png","news",{"type":14,"children":15,"toc":266},"root",[16,24,29,34,43,48,53,61,66,71,78,83,88,95,100,105,112,117,122,127,134,139,144,151,156,161,169,174,179,184,191,196],{"type":17,"tag":18,"props":19,"children":21},"element","h1",{"id":20},"任意pde的秒级求解昇思mindspore最新成果pdeformer-1迎来开源",[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],{"type":23,"value":33},"在包含三百万条不同形式的一维偏微分方程（PDE）数据的广泛预训练后，PDEformer-1模型在训练数据分布内的Zero-shot预测精度，已超越那些仅针对特定方程进行精细训练的专家模型，如FNO和DeepONet。值得一提的是，在面对训练集之外的数据分布时，PDEformer-1更展现出卓越的小样本学习能力，能够凭借少量样本迅速适应并泛化至新的方程，并能够很好地支撑下游任务（如各类反问题）。为了促进学术交流和工业应用，代码及模型参数已经在昇思MindSpore社区开源（链接附在文末），同时相关论文已经被ICLR workshop on AI4DifferentialEquations in Science录用。",{"type":17,"tag":25,"props":35,"children":36},{},[37],{"type":17,"tag":38,"props":39,"children":40},"strong",{},[41],{"type":23,"value":42},"偏微分方程基础模型PDEformer-1，旨在实现不同PDE的通用求解",{"type":17,"tag":25,"props":44,"children":45},{},[46],{"type":23,"value":47},"偏微分方程（PDE）与众多物理现象及工程应用紧密相连，涵盖机翼设计、电磁场模拟、应力分析等广泛领域。在这些实际应用中，PDE的求解往往需要反复进行。PDE数值求解领域当前可分为传统的数值求解方法与AI求解方法两大类别。无论何种方法，面对复杂多变的实际场景，都没有较强的通用性，需要耗费大量的人力和算力来适应新的场景。",{"type":17,"tag":25,"props":49,"children":50},{},[51],{"type":23,"value":52},"华为AI4SCI Lab与北京大学董彬教授团队携手，基于全场景AI框架昇思MindSpore共同提出了PDEformer-1模型，这是一种创新的数据驱动方法。PDEformer-1的核心在于将不同PDE的符号信息转换为多样化的计算图表示形式，然后模型采用Graph Transformer来整合图中的符号和数值信息，最后模型使用隐式神经表示（INR）方案来表达方程各分量的解。值得一提的是，PDEformer-1的构架可以更加灵活的拓展到2D和3D，可实现1D、2D、3D的联合训练。",{"type":17,"tag":25,"props":54,"children":55},{},[56],{"type":17,"tag":57,"props":58,"children":60},"img",{"alt":7,"src":59},"https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2024/04/19/35ad6ae4f44a4aedb22c9d9abf467616.png",[],{"type":17,"tag":25,"props":62,"children":63},{},[64],{"type":23,"value":65},"图1 PDEformer-1架构图（以周期边界对流方程为例）",{"type":17,"tag":25,"props":67,"children":68},{},[69],{"type":23,"value":70},"在预训练阶段，PDEformer-1模型在由300万个样本构成的训练集上取得了0.0084的相对误差表现，而在测试数据集的相对误差为0.0133。下图展示了预训练后的PDEformer-1在测试数据集上的预测结果，该模型在不同PDE问题的求解上均展现出了高度的准确性。",{"type":17,"tag":25,"props":72,"children":73},{},[74],{"type":17,"tag":57,"props":75,"children":77},{"alt":7,"src":76},"https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2024/04/19/b657fc9d34fe4e3ca5a0ef74caf56462.png",[],{"type":17,"tag":25,"props":79,"children":80},{},[81],{"type":23,"value":82},"图2 PDEformer-1在测试数据集上的推理效果（横轴为时间分量，纵轴为空间分量），Reference为传统求解器提供的方程解，Predict为PDEformer-1的预测结果",{"type":17,"tag":25,"props":84,"children":85},{},[86],{"type":23,"value":87},"为了评估预训练后的PDEformer-1在正问题求解方面的性能，研究团队选用了业界标准的PDEBench数据集进行测试，并将所得结果与针对特定PDE设计的专家模型（如FNO和DeepONet）进行了详尽比较。如表1（左）所示，在方程系数处于预训练数据集系数分布范围内（即In-Distribution）的情况下，PDEformer-1展现出了卓越的Zero-shot推理性能，其精度超越了专家模型。表1（右）则展示了Out-of-Distribution (OoD)的结果，PDEformer-1在特定PDEBench数据集上进行微调后（后缀为Finetune）可以达到最低的相对误差。",{"type":17,"tag":25,"props":89,"children":90},{},[91],{"type":17,"tag":57,"props":92,"children":94},{"alt":7,"src":93},"https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2024/04/19/b9a367a5a7ed43a28d17eb2d0e99461c.png",[],{"type":17,"tag":25,"props":96,"children":97},{},[98],{"type":23,"value":99},"表1 不同模型相对误差的对比（左表为系数In-Distribution,右表为OoD）",{"type":17,"tag":25,"props":101,"children":102},{},[103],{"type":23,"value":104},"为了对比微调PDEformer-1与训练专家模型在效率上的差异，研究团队针对PDEformer-1在PDEBench数据集上的OoD情况（即Advection方程，β=1并且仅使用100个训练样本）进行了深入分析。如图3所示，PDEformer-1在微调过程中仅需100次迭代即可达到较高精度。相较之下，从零开始训练的FNO模型则需要数千次迭代才能收敛，且其测试误差也更高。这一对比结果充分展现了PDEformer-1在OoD情况下的快速适应能力和高效性。",{"type":17,"tag":25,"props":106,"children":107},{},[108],{"type":17,"tag":57,"props":109,"children":111},{"alt":7,"src":110},"https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2024/04/19/e4aad9ca29bc426b86c618f6c78b408a.png",[],{"type":17,"tag":25,"props":113,"children":114},{},[115],{"type":23,"value":116},"图3 PDEformer-1微调速度与FNO训练速度比较（横轴为迭代次数，纵轴为相对误差）",{"type":17,"tag":25,"props":118,"children":119},{},[120],{"type":23,"value":121},"PDEformer-1不仅在正问题求解中表现出色，而且也能有效应用于反问题的求解。作为正问题算子的代理模型，它可用于常数系数反演、源项反演以及波方程速度场反演等多种任务，展现出其强大的通用性和实用性。",{"type":17,"tag":25,"props":123,"children":124},{},[125],{"type":23,"value":126},"图4展示了研究团队将粒子群优化算法与PDEformer-1模型相结合，成功应用于常数系数反演任务。图中左侧，Reference代表真实方程解，Noisy为含噪声的方程解，而PDEformer-1实际可获得的观测只有含噪声解的一小部分（如Observed所示）。散点图直观展示了反演得到的方程系数与真实系数之间的高度吻合，几乎所有散点都紧密分布在y=x线上，充分证明了该方法的准确性和有效性。",{"type":17,"tag":25,"props":128,"children":129},{},[130],{"type":17,"tag":57,"props":131,"children":133},{"alt":7,"src":132},"https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2024/04/19/b503b74e33c54cd6abbb1ffbae68ae9d.png",[],{"type":17,"tag":25,"props":135,"children":136},{},[137],{"type":23,"value":138},"图 4 常数系数反演（最右侧散点图中的横、纵坐标分别表示真实值和反演结果）",{"type":17,"tag":25,"props":140,"children":141},{},[142],{"type":23,"value":143},"预训练后的PDEformer-1还可应对更具挑战性的反问题，如源项反演任务。研究团队将源项（即空间依赖的函数）作为待优化向量输入PDEformer-1，并运用梯度下降法进行优化，使模型输出与不完整观测数据吻合。图5展示了PDEformer-1在不同噪声水平下的反演结果，即便在高噪声干扰下，其反演结果与实际源项的偏差依然极小。这一结果充分证明了PDEformer-1在解决复杂反问题时的卓越表现和鲁棒性。",{"type":17,"tag":25,"props":145,"children":146},{},[147],{"type":17,"tag":57,"props":148,"children":150},{"alt":7,"src":149},"https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2024/04/19/922410555569435eb2cfd1386f193021.png",[],{"type":17,"tag":25,"props":152,"children":153},{},[154],{"type":23,"value":155},"图 5 不同噪声水平下的源项反演结果（红色曲线表示反演结果，蓝色曲线为真实的源项）",{"type":17,"tag":25,"props":157,"children":158},{},[159],{"type":23,"value":160},"PDEformer-1在求解PDE正问题与反问题上展现了卓越的灵活性和效率。在PDEBench数据集与专家模型的对比中，PDEformer-1不仅在Zero-shot预测任务中超越了专家模型，更在小样本学习及新任务适应方面凸显出显著优势。同时，在处理如常数系数反演、源项反演等复杂挑战性反问题时，该模型亦展现出高精度和实用性，充分印证了其强大的求解能力。",{"type":17,"tag":25,"props":162,"children":163},{},[164],{"type":17,"tag":38,"props":165,"children":166},{},[167],{"type":23,"value":168},"昇腾AI基础软硬件平台、昇思MindSpore AI框架，支持PDE基础模型探索",{"type":17,"tag":25,"props":170,"children":171},{},[172],{"type":23,"value":173},"依托昇思MindSpore框架的分布式并行能力，以及64块昇腾NPU的强大算力支撑，参数量达到58M的PDEformer-1模型在包含三百万样本的庞大数据集上进行预训练，仅需短短3天即可完成。这一高效表现得益于昇思MindSpore在昇腾AI基础软硬件平台与昇腾CANN的深度集成，通过深度协同优化的高性能算子库，实现了硬件算力的最大化释放。",{"type":17,"tag":25,"props":175,"children":176},{},[177],{"type":23,"value":178},"为了将该PDE基础模型应用于真实场景，研究团队将致力于提升其训练效率，使其能够解决二维、三维乃至更高维度的方程，并研究如何同时利用这些不同维度的数据进行训练。未来不久，研究团队将陆续推出针对二维方程的PDEformer-2、针对三维方程的PDEformer-3，以及融合了不同维度方程的PDEformer。同时，研究团队将扩大考察的方程类型范围，并深入探索更多种类的下游任务。除了已初步研究的反问题，未来还将把PDEformer应用于优化设计、控制、不确定性量化等更广泛的领域。PDEformer能够针对任意类型的PDE，在无网格情况下迅速给出精度可接受的方程解，这让它有望解决传统PDE数值求解方法中编程困难、高计算资源消耗的问题，为所有类型PDE的数值求解提供一种新的范式。",{"type":17,"tag":25,"props":180,"children":181},{},[182],{"type":23,"value":183},"了解更多可查看论文：",{"type":17,"tag":25,"props":185,"children":186},{},[187],{"type":17,"tag":57,"props":188,"children":190},{"alt":7,"src":189},"https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2024/04/19/93398a07c58442e5b146794b8ac3717f.png",[],{"type":17,"tag":25,"props":192,"children":193},{},[194],{"type":23,"value":195},"相关链接：",{"type":17,"tag":197,"props":198,"children":199},"ul",{},[200,218,229,240,254,259],{"type":17,"tag":201,"props":202,"children":203},"li",{},[204,206,210],{"type":23,"value":205},"论文链接：",{"type":17,"tag":207,"props":208,"children":209},"br",{},[],{"type":17,"tag":211,"props":212,"children":216},"a",{"href":213,"rel":214},"https://arxiv.org/abs/2402.12652",[215],"nofollow",[217],{"type":23,"value":213},{"type":17,"tag":201,"props":219,"children":220},{},[221,223],{"type":23,"value":222},"开源代码链接：",{"type":17,"tag":211,"props":224,"children":227},{"href":225,"rel":226},"https://gitee.com/mindspore/mindscience/tree/master/MindFlow/applications/pdeformer1d",[215],[228],{"type":23,"value":225},{"type":17,"tag":201,"props":230,"children":231},{},[232,234],{"type":23,"value":233},"开源数据集链接：",{"type":17,"tag":211,"props":235,"children":238},{"href":236,"rel":237},"https://disk.pku.edu.cn/link/AA6ABF7FEB034446069108D0B6B3920C35",[215],[239],{"type":23,"value":236},{"type":17,"tag":201,"props":241,"children":242},{},[243,245,248],{"type":23,"value":244},"演示视频链接：",{"type":17,"tag":207,"props":246,"children":247},{},[],{"type":17,"tag":211,"props":249,"children":252},{"href":250,"rel":251},"https://b23.tv/2aZW3Hr",[215],[253],{"type":23,"value":250},{"type":17,"tag":201,"props":255,"children":256},{},[257],{"type":23,"value":258},"添加小猫子@MindSpore助手加入微信群或QQ交流群：935025744",{"type":17,"tag":201,"props":260,"children":261},{},[262],{"type":17,"tag":57,"props":263,"children":265},{"alt":7,"src":264},"https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2024/04/19/7918eb0d34334a6986b2a52bd1d555ad.png",[],{"title":7,"searchDepth":267,"depth":267,"links":268},4,[],"markdown","content:news:zh:3070.md","content","news/zh/3070.md","news/zh/3070","md",1776506079295]