AI-Enhanced Signal Integrity Assessment for High-Frequency Chipsets in Next-Generation Wireless Systems

Abstract

Next-generation wireless systems in 6G and beyond foresee increasing interest in chipsets with higher frequencies, including up to WLAN frequencies of 60 GHz and below. Consequently, to maintain signal integrity for these high-frequency chipsets, there are stringent requirements on the PCB layout as well as the antenna-mask fabrication and placement on the PCB. While technological progress has been made on the chip design side and the PCB fabrication side, the signal integrity assessment of the 3D and heterogeneous assembly process is still mostly based on traditional EM simulations, lacking efficiency and intelligence. Instead, a completely new paradigm with AI-enhanced signal integrity assessment is introduced. Benefiting from modern advanced high-performance computing, a 5G chip assembly with ultra-accuracy requirements on signal integrity is studied with a mixed-fidelity approach. Due to its length, the typical hotspots are unseen by exhaustive simulation. Hence, deep learning techniques are employed to accurately identify potential hotspots very fast. An antenna-electrical coupling case study with ultra-lightweight AI architectures shows promising performance on accuracy and speedup.
A comprehensive DQA framework is proposed to assist the AI wireless community in analyzing the DQA problems. Under widely used DQA criteria in the AI model training phase, the proposed framework consists of three tasks to measure the quality of the wireless air-interface data in terms of performance DQA metrics, whereas the effectiveness of the proposed framework is confirmed by validating it on two wireless air-interface data sets. Due to the fundamental limitation in the quality metrics of the synchronous analog data, a task-specific DQA framework is thoroughly benchmarked to fill this gap. With this proposed AI-enabled DQA framework, the DQA of diverse wireless air-interface data can be assessed. Efforts in smoothing the edges of ideal IQ maps and curating additional noise on the existing IQ maps are shown to enhance the robustness of NN models.