A Multiphysics-Based Machine Learning Framework for the Electromagnetic and Thermal Analysis of Electronic Devices

There is plethora of upcoming wireless sensors and Internet of Things (IoT) devices in support of health, communication links, and automotive applications to mention a few. The incorporation of artificial intel-ligence (AI) and machine learning (ML) techniques provides opportunities to introduce more functionality and efficiency for these devices. A typical wireless sensing device may include memory, ASICs, and FPGAs, along with active and passive components. Optimizing the design of these sensors and/or communication devices is important as there is a strong need to reduce their size and weight as well as lower their power consumption and enhance their performance. Currently, these devices are modeled and analyzed using advanced computational tools based on well-established numerical methods. These methods seamlessly integrate electromagnetic (EM), mechanical, and thermal analysis for optimizing device performance. For example, there is a trend for heterogeneous integration of digital and high-power RF packages with adjacent active and passive components. This packaging approach necessitates optimization to mitigate electromagnetic interference (EMI) and ensure electromagnetic compatibility (EMC) in compliance with regulatory standards. While the above conventional approaches have been refined since the 1980s, more recently neural networks have been used to solve differential and integro-differential equations. One such approach is the Physics-Informed Neural Networks (PINNs) method, originally introduced in the 1990s (e.g., Psichogios and Ungar, AIChe, 38, 10, 1493–1682, 1992; Lagaris et al., IEEE trans. Neural Netw., 9, 987–1000, 1998), and later refined by Karniadakis' group, who also coined the term (Rassi et al., Journal of Coup. Phys., 378, 686-707, 2019). PINNs have been extensively studied in fluid mechanics, solid mechanics, chemistry, and geoscience. To our knowledge, our team was the first to introduce an application of PINNs to accurately compute waveguide field distributions (Khan et al., IEEE Access, 12, 120777–120787, 2024). Building on this work, this paper extends the method to incorporate a Multiphysics-based deep-learning framework. The latter integrates two distinct PINN architectures, one dedicated to EM analysis and the other to thermal analysis of electronic devices. Figure 1 provides an overview of our approach. At the conference, we will present the method along with promising results, demonstrating the accuracy and effectiveness of the approach to computing EM and thermal field distributions in complex electronic devices.