CAREER: Neural Network Enhanced Electromagnetics and Multiphysics Simulation Methods for RF and Microwave Reconfigurable Devices

通信、传感和导航系统的快速发展正在推动下一代可重构射频（RF）和微波设备的技术进步。这些设备利用可调的外部刺激，如偏置电压或电流、电/磁/光激励、温度变化和机械力来重新配置其工作特性，并实现频谱敏捷操作。为了解决空间、频谱和功率的限制，小型化和高能效的射频和微波器件的需求很高。虽然可重构性和可控性提供了前所未有的系统灵活性和可靠性，但这些器件的设计和优化方法面临着结构和材料复杂性、多尺度设计挑战、多物理场和非线性相互作用以及高优化维度的巨大挑战。本研究旨在发展物理和神经网络支持的电磁（EM）和多物理场仿真方法，以解决多尺度、多物理场和非线性建模的挑战，从而有效地评估和优化射频和微波可重构器件。该项目着眼于如何开发建模和仿真方法，利用先进的数值和神经网络技术，实现更高效、更可靠的设备建模和评估。此外，该项目还制定了广泛的教育和外联计划，包括让非洲裔美国学生和其他代表性不足的少数民族学生参与其中，以及制作视频剪辑和示范，向公众传播结果。为了开发、实施、验证和应用电磁和射频/微波可重构器件多物理场设计的建模和仿真方法，该项目将开展四项主要研究活动：1)将开发一种新的全频率稳定电磁公式及其域分解方法（DDM），以解决宽带和多尺度电磁建模问题。2)利用图神经网络（graph neural network， GNN）辅助的DDM将以优越的效率解决大规模EM问题。3)研究神经网络（NN）辅助多物理场仿真方法和非线性替代求解器，以解决多物理场建模和求解非线性问题的挑战，而无需传统的基于梯度或牛顿的迭代。4)物理引导的神经网络设备优化器将集成上述数值评估技术，提供快速的参数扫描和形状优化能力，以对抗高维数。将物理和神经网络科学计算方法无缝集成的研究将导致一种革命性的仿真工具，具有目前不存在的建模和设计能力。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

The rapid development of communication, sensing, and navigation systems are driving technological advances in the next-generation reconfigurable radiofrequency (RF) and microwave devices. These devices utilize tunable external stimuli such as biasing voltages or currents, electrical/magnetic/optical excitations, temperature variations, and mechanical forces to reconfigure their working properties and achieve spectrum-agile operations. To address the spatial, spectrum, and power limits, miniaturized and power-efficient RF and microwave devices are of high demand. While the reconfigurability and controllability provide unprecedented system flexibility and reliability, the design and optimization methods for such devices face great challenges coming from the structural and material complexities, multiscale design challenges, multiphysics and nonlinear interactions, and high optimization dimensionalities. This research aims at developing physics and neural network enabled electromagnetic (EM) and multiphysics simulation methods to address the challenges of multiscale, multiphysics, and nonlinear modeling for the efficient evaluation and optimization of RF and microwave reconfigurable devices. The project looks at how to develop modeling and simulation methods that utilize advanced numerical and neural network techniques for more efficient and reliable device modeling and assessments. In addition, the project has extensive education and outreach plans including the involvement of African American students and other underrepresented minority students, as well as the development of video clips and demonstrations to disseminate the results to the public. To develop, implement, validate, and apply modeling and simulation methods for EM and multiphysics design of RF/microwave reconfigurable devices, the project will conduct four major research activities: 1) A novel all-frequency stable EM formulation and its domain-decomposition method (DDM) will be developed to address wideband and multiscale EM modeling problems. 2) A graph neural network (GNN)-aided DDM will be developed to solve large-scale EM problems with a superior efficiency. 3) Neural network (NN)-assisted multiphysics simulation method and nonlinear surrogate solvers will be investigated to address challenges in multiphysics modeling and solve nonlinear problems without the need of traditional gradient- or Newton-based iteration. 4) A physics-guided NN device optimizer will integrate the above numerical evaluation techniques to provide fast parameter sweep and shape optimization capabilities to combat the high dimensionality. The research to seamlessly integrate physics- and NN-enabled scientific computing methodologies will lead to a revolutionary simulation tool with enabling modeling and design capabilities that do not currently exist.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.