diff --git a/sigs/mindelec/README.md b/sigs/mindelec/README.md new file mode 100644 index 0000000..78a91a9 --- /dev/null +++ b/sigs/mindelec/README.md @@ -0,0 +1,72 @@ +## SIG简介 + +电磁场看不见、摸不着，却在日常生活中无处不在。电磁场的产生主要源于自然和人工两类。自然产生的电磁场有地磁场、太阳光以及一切物体热辐射产生的电磁波等等。自然电磁场催生并推动了人类的文明：由于太阳光的存在，人类可以在温度适宜的地球居住，可以通过植物的光合作用获取充足的食物；利用地磁场人类可以进行导航，进而迎来了大航海和全球化时代。随着科技的发展，人类已经不满足于自然产生的电磁场，开始主动向环境中发射电磁场，并充分挖掘电磁场的应用潜力。如通信领域中利用无线电波收听广播、利用高频微波进行手机通话等；又如地质勘探中利用电磁波的回波探明煤炭存储等。 + +电磁场的应用不胜枚举，为了能够更好的利用电磁场，人们通过实验、理论以及计算等手段研究电磁场的机理。实验方面，1820年奥斯特在一次讲座上偶然发现通电的导线让小磁针发生偏转，从而发现了电能生磁的现象。1831年，法拉第在实验中发现变化的磁场可以产生电场，即磁也能够生电。麦克斯韦总结前人的工作，提出了位移电流假说（变化的电场能够产生磁场），完善了电生磁的理论。最终，麦克斯韦将电磁场理论用给简洁、对称和完美的数学形式表示出来，即麦克斯韦方程组。随着计算机技术的发展，人们采用数值计算的方式去求解麦克斯韦方程组,模拟电磁场在空间中的分布。这样即可以节省实验的成本，也可以通过仿真设计出更加符合需求的电子设备。传统的电磁计算方法包括精确的全波方法和高频近似方法。全波方法如时域有限差分（Finite-Difference Time-Domain method，FDTD），有限元（Finite-Element-Method，FEM）、矩量法（Method of MoMents，MoM）等；高频近似方法如几何光学法（Geometrical Optics，GO）、物理光学法(physical optics，PO)等。 + +数值计算较好地辅助了电子产品的设计，但传统的数值方法仍存在许多缺陷，如需要进行复杂的网格剖分、迭代计算，计算过程复杂、计算周期长。神经网络具有万能逼近和高效推理能力，这使得神经网络在求解微分方程时具有潜在的优势。为此，昇思MindSpore Elec AI 专项兴趣小组（简称：智能电磁AI SIG）正式成立，并面向开源社区招募志同道合的伙伴。 + +## MindSpore Elec AI SIG的使命 + +围绕实际生产中的各类电磁应用场景，在昇思MindSpore框架下探索和研究基于AI的电磁正问题以及反问题。例如开发高效精确的AI电磁模型，构建高效的MindSpore Elec基础工具包，提升电子产品的设计效率等。 + +### MindSpore Elec 代码仓 + +1. [Mindspore Elec 代码仓](https://gitee.com/mindspore/mindscience/tree/master/MindElec) +2. [Mindspore Elec SIG 工作目录]( https://gitee.com/mindspore/community/tree/master/sigs/mindelec) + +### SIG小组重点工作方向 + +#### MindSpore Elec基础工具包构建 + +基于MindSpore构建MindSpore Elec基础工具包。基础工具包内置有数据构建及转换、仿真计算以及结果可视化等。 + + 1. 数据构建及转换：支持CSG （Constructive Solid Geometry，CSG） 模式的几何构建，以及cst和stp数据（CST等商业软件支持的数据格式）的高效张量转换。 + + 2. 仿真计算： + a) AI电磁基础模型库：提供物理和数据驱动的AI电磁模型，物理驱动无需额外的标签数据，只需方程和初边界条件即可；数据驱动是指训练需使用仿真或实验等产生的数据。 + b) 优化策略：数据压缩可以有效地减少神经网络的存储和计算量；多尺度滤波、动态自适应加权可以提升模型的精度，克服点源奇异性等问题；小样本学习主要是为了减少训练的数据量，节省训练的成本。 + + 3. 结果可视化：仿真的结果如S参数或电磁场等可保存在CSV、VTK文件中。MindInsight可以显示训练过程中的损失函数变化，并以图片的形式在网页上展示结果；Paraview是第三方开源软件，具有动态展示切片、翻转等高级功能。 + +#### AI电磁仿真模型和方法构建 + + 1. 端到端可微的传统电磁仿真方法研究：基于MindSpore构筑传统的电磁仿真方法如FDTD/有限元/矩量法等，形成端到端可微的AI融合方法。这样可以利用MindSpore加速传统的数值方法，生成模型训练的数据，也可以基于自动微分机制，实现数据同化、电磁反演等应用。 + + 2. AI电磁仿真融合算法研究: 物理驱动（如PINNs方法）和数据驱动的AI方法，以及物理和数据融合的算法创新等。 + +#### AI电磁仿真模型应用 + + 1. 正问题：基站天线、雷达天线、射频电路与系统等电磁仿真。 + + 2. 反问题：电磁超材料设计、雷达勘探、电磁成像等。 + +## MindSpore Elec AI SIG工作计划 + + 1. 前期：以成员学术交流活动为主，每月组织线上交流活动，围绕AI电磁中涉及的问题，介绍研究工作进展，讨论研究工作中的难点。 + + 2. 后期：通过合作开发等模式，在国内高校及企业间开展电磁AI问题的合作研究。 + +## MindSpore Elec AI SIG构成 + +### 领衔成员： + +* 陆卫兵 东南大学科研院 院长/教授 +* 杨武 东南大学信息科学与工程学院 副研究员 +* 俞文明 东南大学信息科学与工程学院 教授 +* 徐勇 江苏浩云连德信息技术有限公司 高级工程师 +* 李家奇 东南大学物理学院 副教授 + +### 小组成员： + +* 成员：苑玉杰 华为昇思MindSpore布道师 +* 成员：Kyang 华为昇思MindSpore高级工程师 +* 成员：刘彻 东南大学信息科学与工程学院 博士后 +* 成员：鲍江涵 东南大学信息科学与工程学院 博士 + +* 成员：翁瑞 东南大学信息科学与工程学院 博士生 +* 成员：Adrian Lee 华为昇思MindSpore高级工程师 +* 成员：Lulu Zhang 华为昇思MindSpore高级工程师 +* 成员：秦洁 东南大学信息科学与工程学院 硕士生 +* 成员：张哲 东南大学信息科学与工程学院 硕士生 +* 成员：孙丁一 东南大学信息科学与工程学院 硕士生 diff --git a/sigs/mindelec/README_en.md b/sigs/mindelec/README_en.md new file mode 100644 index 0000000..6b431fb --- /dev/null +++ b/sigs/mindelec/README_en.md @@ -0,0 +1,72 @@ +## Introduction to SIG + +Electromagnetic fields cannot be seen or touched, but they are ubiquitous in daily life. Electromagnetic fields are mainly generated by two ways: natural and artificial. Natural electromagnetic fields include the earth's magnetic field, sunlight, and electromagnetic waves generated by thermal radiation of all objects, etc. The natural electromagnetic field gave birth to and promoted human civilization: due to the existence of sunlight, human beings can live on the earth with a suitable temperature, and can obtain sufficient food through the photosynthesis of plants; human beings can navigate by using the geomagnetic field, and then ushered in great voyages and the age of globalization. With the development of science and technology, human beings are no longer satisfied with the natural electromagnetic fields, and have begun to actively emit electromagnetic fields into the environment, and fully tap the application potential of electromagnetic fields. For example, in the field of communication, radio waves are used to listen to broadcasts, high-frequency microwaves are used for mobile phone calls, etc.; another example is the use of electromagnetic wave echoes to prove coal storage in geological exploration. + +The applications of electromagnetic fields are numerous. In order to make better use of electromagnetic fields, people study the mechanism of electromagnetic fields through experiments, theories, and calculations. In terms of experiments, in 1820, Oersted accidentally discovered that an energized wire deflected a small magnetic needle in a lecture, thus discovering the phenomenon of electric energy generating magnetism. In 1831, Faraday discovered in experiments that a changing magnetic field can generate an electric field, that is, magnetism can also generate electricity. Maxwell summed up the work of predecessors, put forward the hypothesis of displacement current (changing electric field can generate magnetic field), and perfected the theory of electromagnetism. Ultimately, Maxwell expressed the electromagnetic field theory in a concise, symmetrical and perfect mathematical form, namely Maxwell's equations. With the development of computer technology, people use numerical calculation methods to solve Maxwell's equations and simulate the distribution of electromagnetic fields in space. In this way, the cost of experiments can be saved, and electronic devices that better meet the needs can be designed through simulation. Traditional electromagnetic calculation methods include accurate full-wave methods and high-frequency approximation methods. Full-wave methods such as Finite-Difference Time-Domain method (FDTD), finite element (Finite-Element-Method, FEM), method of moments (Method of MoMents, MoM), etc.; high-frequency approximation methods such as geometric Optical method (Geometrical Optics, GO), physical optics (physical optics, PO), etc. + +Numerical calculations can better assist the design of electronic products, but there are still many defects in traditional numerical methods, such as the need for complex grid division, iterative calculations, complex calculation processes, and long calculation cycles. The universal approximation and high-efficiency reasoning capabilities of neural networks give neural networks potential advantages in solving differential equations. To this end, Shengsi MindSpore Elec AI Special Interest Group (abbreviation: Intelligent Electromagnetic AI SIG) was formally established, and is recruiting like-minded partners from the open source community. + +## MindSpore Elec AI SIG Mission + +Focusing on various electromagnetic application scenarios in actual production, explore and study AI-based electromagnetic forward and inverse problems under the framework of Shengsi MindSpore. For example, develop an efficient and accurate AI electromagnetic model, build an efficient MindSpore Elec basic toolkit, and improve the design efficiency of electronic products. + +### MindSpore Elec code repository + +1. [Mindspore Elec repository](https://gitee.com/mindspore/mindscience/tree/master/MindElec) +2. [Mindspore Elec SIG repository]( https://gitee.com/mindspore/community/tree/master/sigs/mindelec) + +### SIG group key work direction + +#### MindSpore Elec basic toolkit construction + +Build the MindSpore Elec basic toolkit based on MindSpore. The basic toolkit has built-in data construction and conversion, simulation calculation and result visualization, etc. + + 1. Data construction and conversion: support the geometric construction of CSG (Constructive Solid Geometry, CSG) mode, and the efficient tensor conversion of cst and stp data (data formats supported by commercial software such as CST). + + 2. Simulation calculation： + a) AI electromagnetic basic model library: Provides physical and data-driven AI electromagnetic models. Physical driving does not require additional label data, only equations and initial boundary conditions are required; data-driven means that training needs to use data generated by simulation or experiments. + b) Optimization strategy: data compression can effectively reduce the amount of storage and calculation of the neural network; multi-scale filtering and dynamic adaptive weighting can improve the accuracy of the model and overcome problems such as point source singularity; small sample learning is mainly to reduce the amount of training data , saving training costs. + + 3. Result visualization: Simulation results such as S parameters or electromagnetic fields can be saved in CSV and VTK files. MindInsight can display the changes in the loss function during the training process, and display the results on the webpage in the form of pictures; Paraview is a third-party open source software that has advanced functions such as dynamically displaying slices and flips. + +#### AI electromagnetic simulation model and method construction + + 1. Research on end-to-end differentiable traditional electromagnetic simulation methods: build traditional electromagnetic simulation methods such as FDTD/finite element/moment method based on MindSpore, and form an end-to-end differentiable AI fusion method. In this way, MindSpore can be used to accelerate traditional numerical methods to generate data for model training, and it can also implement applications such as data assimilation and electromagnetic inversion based on the automatic differentiation mechanism. + + 2. AI electromagnetic simulation fusion algorithm research: physics-driven (such as PINNs method) and data-driven AI methods, and algorithm innovation for physics and data fusion, etc. + +#### Application of AI electromagnetic simulation model + + 1. Positive problems: Electromagnetic simulation of base station antennas, radar antennas, radio frequency circuits and systems. + + 2. Inverse problems: electromagnetic metamaterial design, radar exploration, electromagnetic imaging, etc. + +## MindSpore Elec AI SIG work plan + + 1. Early stage: Focusing on members' academic exchange activities, monthly online exchange activities are organized, focusing on the issues involved in AI electromagnetics, introducing the research progress and discussing the difficulties in the research work. + + 2. Later stage: Through cooperative development and other modes, carry out cooperative research on electromagnetic AI issues among domestic universities and enterprises. + +## Composition of MindSpore Elec AI SIG + +### lead members： + +* Weibing Lu, Dean/Professor, Research Institute of Southeast University +* Wu Yang, Associate Researcher, School of Information Science and Engineering, Southeast University +* Wenming Yu, Professor, School of Information Science and Engineering, Southeast University +* Wenming Yu, Professor, School of Information Science and Engineering, Southeast University +* Jiaqi Li, Associate Professor, School of Physics, Southeast University + +### team members： + +* Members: Yujie Yuan, Shengsi MindSpore evangelist +* Member: Kyang, Senior Engineer of Huawei MindSpore +* Member: Che Liu, Postdoctoral Fellow, School of Information Science and Engineering, Southeast University +* Member: Jianghan Bao, Ph.D., School of Information Science and Engineering, Southeast University + +* Member: Rui Weng, Ph.D student, School of Information Science and Engineering, Southeast University +* Member: Adrian Lee, Senior Engineer of Huawei MindSpore +* Member: Lulu Zhang, Senior Engineer of Huawei MindSpore +* Member: Jie Qin, Master student, School of Information Science and Engineering, Southeast University +* Member: Zhe Zhang, Master student, School of Information Science and Engineering, Southeast University +* Member: Dingyi Sun, Master student, School of Information Science and Engineering, Southeast University