MS004321-人工智能与电磁学协同赋能技术

研究生课程开设申请表

开课院（系、所）：3044am永利集团3044noc      

课程申请开设类型： 新开     重开□     更名□（请在□内打勾，下同）

一、课程介绍（含教学目标、教学要求等）（300字以内）

《人工智能与电磁学协同赋能技术》面向5G/6G、雷达与智能感知场景，讲授AI与电磁互促的最新理论、方法及实践。课程分两条主线：AI加速毫米波器件、天线、超表面建模、逆设计与优化；可编程超材料、衍射神经网与片上电磁算子为AI提供低功耗硬件，实现“物理即计算”。内容涵盖电磁-AI基础、智能仿真、可重构超表面、衍射神经网络、光电融合芯片及具身多模态交互。以理论-算法-实验-系统四层次培养跨学科创新力，并通过“国产毫米波芯片突围”等案例融入思政，培育具有全球视野与家国情怀的信息人才。

二、教学大纲（含章节目录）：（可附页）

1. 绪论和智能计算电磁学的算法篇：（4学时）

智能电磁计算在毫米波通信与雷达需求激增的背景下应运而生，融合电磁理论、人工智能与高性能计算三大要素。本课程将系统阐述其发展脉络，并指出毫米波频段成为验证新方法论的关键战场。回顾传统计算电磁学，涵盖微分形式、积分形式及半解析三类数值算法，为后续智能化改造奠定离散与逼近基础。随后转向人工智能科学计算，先剖析前馈、循环、深度与生成式神经网络的电磁建模潜力，再讨论遗传、粒子群、差分进化与空间映射等优化算法在逆设计和调参中的应用。通过将物理约束嵌入可微分网络并与优化器协同，可实现精度与效率的指数级提升。上述内容共同勾勒出“数据-物理-算法”三元融合的智能电磁计算完整技术路线。

2. 智能计算电磁学的数据篇：（4学时）

本次课程将介绍三种数据驱动框架：纯数据拟合的正向计算、端到端映射的逆向计算，以及正逆闭环自监督的融合计算。数据驱动智能计算方法包括正向计算（基于数据建模电磁响应）、逆向计算（通过数据反演电磁参数）以及正逆结合计算（实现闭环优化），为电磁问题提供高效的数据驱动解决方案。数理驱动智能计算方法分为外挂式（将物理约束作为优化目标）和内嵌式（将物理方程嵌入网络结构），通过融合数学模型提升计算精度和泛化能力。物理驱动智能计算方法则进一步发展出频域和时域物理信息神经网络，将麦克斯韦方程组等物理规律直接编码到神经网络中，实现物理一致的智能计算。这三种方法各具特色，数据驱动方法擅长处理复杂非线性关系，数理驱动方法注重数学模型融合，物理驱动方法则强调物理规律约束。它们共同构成了智能电磁计算的方法体系，推动着电磁仿真与设计向更智能、更高效的方向发展。

3. 智能计算电磁学的算力篇：（4学时）

本次课程从单机多CPU、多GPU到跨机集群，系统梳理了电子计算机在电磁计算中的三级并行加速策略，显著扩展了可解问题的规模与实时性。随后则跳脱传统冯·诺依曼架构，提出空间与平面两种电磁衍射神经网络，以光速完成前向传播与反向调控。前者利用三维自由空间衍射实现高并行度，后者在平面波导阵列上完成片上计算，兼顾集成度与能效。两种“电磁计算机”将物理过程本身转化为计算资源，为毫米波场景带来了数量级的延迟降低与功耗节省。电子并行与电磁原生计算的互补融合，正把智能电磁计算推向实时、高效、低能耗的新阶段。

4. 智能计算电磁学的应用篇：（4学时）

本次课程将介绍智能电磁计算在毫米波电路器件等领域的突破性应用，通过智能算法优化无源器件（如滤波器和耦合器）和有源器件（如放大器和混频器）的设计流程，显著提升了器件性能。在毫米波超材料方面，该方法不仅能够智能设计具有特定电磁特性的超材料单元结构，还能高效优化大规模超材料阵列的排布方式，实现精准的电磁波调控。对于毫米波天线系统，智能计算技术既可用于单个天线的快速优化设计，也能处理复杂天线阵列的综合优化问题，大幅缩短设计周期。这些应用案例充分展示了智能算法在解决高频电磁问题时的独特优势，包括处理复杂非线性关系和实现多目标优化。通过将人工智能与传统电磁设计方法相结合，智能电磁计算正在推动毫米波器件从理论设计到工程应用的全链条创新。

5. 专题实验（6学时）

采用当下主流的编程平台（Matlab、C++或Python）进行编程实验，可由最多3人组成小组完成，一共3次，内容分别为：

1. 数据驱动人工神经网络训练实验

2. 数理驱动人工神经网络训练实验

3. 物理驱动人工神经网络训练实验

要求完成：

1. 完整的编程代码

2. 详细的编程实验报告word文档

6 研讨（10学时）

 针对相关编程实验进行研讨：

1. 数据驱动人工神经网络训练实验（2学时）

2. 数理驱动人工神经网络训练实验（2学时）

3. 物理驱动人工神经网络训练实验（2学时）

针对当下的主流算法和应用进行研讨：

1. 电磁计算机加速计算技术（2学时）

2. 电磁超材料的智能设计（2学时）

要求完成：

1. 研讨ppt文稿

2. 分组演讲并回答研讨提问

三、教学周历

注：1.以上一、二、三项内容将作为中文教学大纲，在研究生院中文网页上公布，四、五内容将保存在研究生院。2.开课学期为：春季、秋季或春秋季。3.授课语言为：汉语、英语或双语教学。4.适用学科范围为：公共，一级，二级，三级。5.实践环节为：实验、调研、研究报告等。6.教学方式为：讲课、讨论、实验等。7.学位课程考试必须是笔试。8.课件地址指在网络上已经有的课程课件地址。9.主讲教师简介主要为基本信息（出生年月、性别、学历学位、专业职称等）、研究方向、教学与科研成果，以100至500字为宜。

四、主讲教师简介：

游检卫，3044am永利集团3044noc教授，博士生导师，3044am永利集团3044noc青年首席教授，国家级高层次青年人才，毫米波全国重点实验室“智能电磁协同赋能技术(AIEM)”团队负责人。长期从事 “计算电磁学”和“电磁超材料”的软硬件开发，相关研究成果已发表100余篇学术论文、专著/章节5部，包括第一作者/通信作者发表Science子刊、Nature子刊、Adv. Funct. Mater.、IEEE Trans.旗舰期刊等。拥有自主知识产权的软硬件平台已服务于我国诸多重点单位，获中国高等学校十大科技进展、中央高校优秀青年团队、江苏省科学技术一等奖等。主持和参与欧盟地平线计划欧洲研究理事会（ERC）重大项目、英国工程与自然科学研究理事会（EPSRC）项目、国家自然科学基金项目、国家基础科学中心项目、国家重点研发计划项目、江苏省自然科学基金项目、南京市留学人员科技创新择优资助项目等。

五、任课教师信息（包括主讲教师）：

六、课程必要性说明

（包括不限于，对比已有课程的重复度？对已有课程的补充性？）

研究生课程《机器学习原理与应用》的教学目标在于让学生掌握机器学习方面的基础理论、常用算法和应用技术，建立机器学习相关的知识结构，提升利用机器学习解决实际问题的能力。授课内容主要包括常见的机器学习算法（线性模型、决策树、神经网络、支持向量机、贝叶斯分类等）的工作原理。

研究生课程《机器学习与进化计算》的教学目标在于使学生掌握多种机器学习范型、算法及进化计算在机器学习领域的研究和应用，吸取包括概念学习、决策树学习、人工神经网络知识、统计和估计理论、贝叶斯观点、计算学习理论、基于实例的学习方法、进化计算理论、学习规则集合的算法、分析学习、归纳与分析学习相结合以及增强学习方面的研究成果。

上述两门课程均侧重于对机器学习算法和进化计算算法工作原理的介绍，缺乏如何利用其在电磁场与微波技术领域，尤其是计算电磁学和电磁超材料专业方向发挥作用的讲解。近年来，随着以深度学习为代表的新一代人工智能的快速发展，人工神经网络目前已被广泛应用于计算电磁学和电磁超材料专业方向，其发展前景得到人们广泛认可。亟需开展一门以融合新一代人工智能的智能电磁学为主要内容的研究生课程，为我国建立从智能电磁计算到智能电磁应用的高水平人才库奠定基础。

七、课程开设审批意见

所在院（系）

审 批 意 见

                 负责人：

                 日  期：

所在学位评定分

委员会审批意见

                      分委员会主席：

                日  期：

研究生院审批意见

                负责人：

                日  期：

  注

说明：1．研究生课程重开、更名申请也采用此表。表格下载：http: /seugs.seu.edu.cn/down/1.asp

2．此表一式三份，交研究生院、院（系）和自留各一份，同时提交电子文档交研究生院。

Application Form For Opening Graduate Courses

School (Department/Institute)：

Course Type: New Open   Reopen □   Rename □（Please tick in □, the same below）

1. Course Introduction (including teaching goals and requirements) within 300 words:

《Bidirectional Empowerment Technology of Artificial Intelligence and Electromagnetics》is designed for 5G/6G, radar and intelligent-sensing scenarios, presenting the latest AI-electromagnetics reciprocity theories, methods, and hands-on practice. The course follows two parallel tracks: (1) AI-accelerated modeling, inverse design and optimization of millimetre-wave devices, antennas and metasurfaces; (2) programmable metamaterials, diffractive neural networks and on-chip electromagnetic operators that deliver ultra-low-power AI hardware, embodying “physics-as-computation.” Topics span electromagnetics-AI fundamentals, intelligent simulation, reconfigurable intelligent surfaces, diffractive neural networks, opto-electronic fusion chips and embodied multimodal interaction. A four-level pedagogy—theory, algorithm, experiment and system—cultivates interdisciplinary innovation, while case studies such as “China’s mm-wave chip breakout” integrate ideological education, nurturing information professionals with global vision and national commitment.

2. Teaching Syllabus (including the content of chapters and sections. A sheet can be attached):

1. Introduction and Algorithmic Foundations of Intelligent Computational Electromagnetics: (4 class hours)

Intelligent computational electromagnetics emerged in response to the explosive demand for millimetre-wave communications and radar, converging electromagnetic theory, artificial intelligence and high-performance computing. This course systematically traces its evolution and pinpoints the millimetre-wave band as the critical arena for validating new methodologies. We first revisit classical computational electromagnetics, covering differential, integral and semi-analytic numerical algorithms to establish the discrete and approximation foundations for subsequent intelligent enhancements. Turning to AI-driven scientific computing, we dissect the electromagnetic modelling potential of feed-forward, recurrent, deep and generative neural networks, then examine genetic algorithms, particle swarm, differential evolution and space mapping for inverse design and parameter tuning. By embedding physical constraints into differentiable networks and orchestrating them with optimisers, accuracy and efficiency can be boosted exponentially. Together, these components outline the complete “data-physics-algorithm” triadic roadmap of intelligent computational electromagnetics.

2. Data of Intelligent Computational Electromagnetics: (4 class hours)

Intelligent computational electromagnetics emerged in response to the explosive demand for millimetre-wave communications and radar, converging electromagnetic theory, artificial intelligence and high-performance computing. This course systematically traces its evolution and pinpoints the millimetre-wave band as the critical arena for validating new methodologies. We first revisit classical computational electromagnetics, covering differential, integral and semi-analytic numerical algorithms to establish the discrete and approximation foundations for subsequent intelligent enhancements. Turning to AI-driven scientific computing, we dissect the electromagnetic modelling potential of feed-forward, recurrent, deep and generative neural networks, then examine genetic algorithms, particle swarm, differential evolution and space mapping for inverse design and parameter tuning. By embedding physical constraints into differentiable networks and orchestrating them with optimisers, accuracy and efficiency can be boosted exponentially. Together, these components outline the complete “data-physics-algorithm” triadic roadmap of intelligent computational electromagnetics.

3. Computing Power for Intelligent Computational Electromagnetics: (4 class hours)

This lecture systematically maps the three-tier parallel-acceleration playbook for electronic computers in electromagnetic computations: from single-node multi-CPU and multi-GPU setups to cross-node clusters, dramatically scaling solvable problem sizes and real-time capability. Stepping beyond the conventional von Neumann paradigm, we then introduce spatial and planar electromagnetic diffractive neural networks that perform forward propagation and backward tuning at the speed of light. The spatial variant harnesses 3-D free-space diffraction for massive parallelism, while the planar counterpart executes on-chip computation in waveguide arrays, balancing integration density with energy efficiency. Both “electromagnetic computers” turn the physical process itself into a computing resource, delivering order-of-magnitude reductions in latency and power for millimetre-wave scenarios. The complementary fusion of electronic parallelism and native electromagnetic computation is propelling intelligent computational electromagnetics into a new era that is real-time, highly efficient, and ultra-low-power.

4. Application of Intelligent Computational Electromagnetics: (4 class hours)

This course will introduce the groundbreaking applications of intelligent computational electromagnetics in fields such as millimeter-wave circuit devices. By leveraging intelligent algorithms to optimize the design processes of passive components (e.g., filters and couplers) and active components (e.g., amplifiers and mixers), the performance of these devices has been significantly enhanced. In the domain of millimeter-wave metamaterials, this approach not only enables the intelligent design of unit structures with specific electromagnetic properties but also efficiently optimizes the arrangement of large-scale metamaterial arrays, achieving precise electromagnetic wave control. For millimeter-wave antenna systems, intelligent computing technology can be applied to both the rapid optimization of individual antennas and the comprehensive optimization of complex antenna arrays, substantially reducing design cycles. These application cases fully demonstrate the unique advantages of intelligent algorithms in addressing high-frequency electromagnetic challenges, including handling complex nonlinear relationships and achieving multi-objective optimization. By integrating artificial intelligence with traditional electromagnetic design methods, intelligent computational electromagnetics is driving end-to-end innovation in millimeter-wave devices—from theoretical design to engineering applications..

5. Special Experiment (6 class hours)

Adopting current mainstream programming platforms (Matlab C++or Python programming experiments can be conducted in groups of up to 3 people, for a total of 3 times. The content includes:

(1) Data-Driven Artificial Neural Network Training Experiment

(2) Mathematics-Driven Artificial Neural Network Training Experiment

(3) Physics-Driven Artificial Neural Network Training Experiment

Required completion:

• Complete programming code

• Detailed programming experiment report Word document

6. Seminars (10 class hours)

Discussion on relevant programming experiments:

(1) Data-Driven Artificial Neural Network Training Experiment (2 class hours)

(2) Mathematics-Driven Artificial Neural Network Training Experiment (2 class hours)

(3) Physics-Driven Artificial Neural Network Training Experiment (2 class hours)

Conduct discussions on current mainstream software/algorithms:

(1) Electromagnetic Computer–Accelerated Computing Techniques (2 class hours)

(2) Intelligent Design of Electromagnetic Metamaterials (2 class hours)

Required completion:

• Discuss the PowerPoint presentation

• Group presentations and answering discussion questions

3. Teaching Schedule:

Note: 1.Above one, two, and three items are used as teaching Syllabus in Chinese and announced on the Chinese website of Graduate School. The four and five items are preserved in Graduate School.

2. Course terms: Spring, Autumn , and Spring-Autumn term.

3. The teaching languages for courses: Chinese, English or Chinese-English.

4. Applicable range of discipline: public, first-class discipline, second-class discipline, and third-class discipline.

5. Practice includes: experiment, investigation, research report, etc.

6. Teaching methods: lecture, seminar, practice, etc.

7. Examination for degree courses must be in paper.

8. Teaching material websites are those which have already been announced.

9. Brief introduction of chief lecturer should include: personal information (date of birth, gender, degree achieved, professional title), research direction, teaching and research achievements. (within 100-500 words)

4. Brief Introduction of Chief lecturer:

Jian Wei You is a Professor and Doctoral Supervisor at Southeast University, where he holds the prestigious title of Young Chief Professor. Recognized as a National High-Level Young Talent, he leads the "AI-Enabled Intelligent Electromagnetic Technologies (AIEM)" team at the National Key Laboratory of Millimeter Waves.  With a long-term focus on the software and hardware development of "Computational Electromagnetics" and "Electromagnetic Metamaterials," his research has resulted in over 100 academic publications, including five books/book chapters. His first-author and corresponding-author works have appeared in prestigious journals such as Science sub-journals, Nature sub-journals, Advanced Functional Materials, and flagship IEEE Transactions journals. His independently developed software and hardware platforms, protected by intellectual property rights, have been adopted by key national institutions. His contributions have earned him accolades such as the "Top 10 Scientific and Technological Advances in Chinese Universities," the "Outstanding Young Team in Central Universities," and the "First Prize of Jiangsu Provincial Science and Technology Award." He has led and participated in major research initiatives, including the European Research Council (ERC) projects under the EU Horizon Programme, the UK Engineering and Physical Sciences Research Council (EPSRC) projects, the National Natural Science Foundation of China (NSFC) projects, the National Basic Science Center projects, the National Key R&D Program projects, the Jiangsu Provincial Natural Science Foundation projects, and the Nanjing Overseas Students Science and Technology Innovation Funding projects..

5. Lecturer Information (include chief lecturer)