Thermal Simulation of Chiplet heterogeneous Integration system

  Prof. Qinzhi Xu

  Research Professor，Institute of Microelectronics of Chinese Academy of Sciences, China    

Abstract:

As chip dimensions continue to scale down, chiplet-based heterogeneous integration has emerged as a pivotal approach for extending Moore’s law, yet it also introduces formidable thermal management challenges. In recent years, physics-informed neural networks (PINNs) have demonstrated promising potential for thermal simulation of integrated circuits; however, when applied to structures with complex geometries, they are prone to becoming trapped in local minima. In this work, we embed numerical computation methods into neural operator learning and propose a hybrid training framework that couples DeepONet with hard-constraint projection. This framework is further extended to the thermal simulation of chiplet heterogeneous integration systems. By incorporating Latin hypercube sampling and grid-based sampling to strengthen physical constraints, high-fidelity temperature simulations of the chiplet system are realized. Numerical experiments confirm that the proposed method yields temperature predictions consistent with conventional numerical simulations while achieving a speedup of nearly 400×, underscoring the potential of physics-guided deep learning for thermal analysis of advanced electronic packaging.

Speaker's Biography:  

Xu Qinzhi, Ph.D. and Professor and Doctoral Supervisor at the Institute of Microelectronics of Chinese Academy of Sciences (CAS). He has long been engaged in research on multiphysics simulation of three-dimensional integrated chips, design-for-manufacturability (DFM) methodologies for nanoscale chips, multiphysics modeling approaches for advanced device processes, structure–property relationships of nanocomposites based on first-principles calculations, and electronic design automation (EDA) software tools. He possesses extensive R&D experience in key simulation models for critical nanoscale chip processes and in EDA technologies for multiphysics simulation of integrated chips. He has proposed a coupled density functional theory (DFT) and integral equation simulation method for multi-site polymer nanocomposites, and developed a series of multiphysics simulation model engines for HKMG, FinFET, and copper-interconnect CMP processes, as well as a multiphysics simulation tool for chiplet integration system. He has undertaken nearly 30 projects funded by national programs, the Beijing municipal government, the Chinese Academy of Sciences, and industry collaboration initiatives, with related technological achievements successfully transferred to well-known enterprises. As the first author or corresponding author, he has published over 100 SCI papers, patents, and software copyrights. He has received awards including the Third Prize of the Beijing Municipal Science and Technology Award and the Second Prize of the Science and Technology Award of the Chinese Institute of Electronics.