edfas.org ELECTRONIC DEVICE FAILURE ANALYSIS | VOLUME 26 NO. 2 10 MICROSTRUCTURAL HIERARCHY DESCRIPTOR ENABLING INTERPRETATIVE AI FOR MICROELECTRONIC FAILURE ANALYSIS Zhiheng Huang1, Ziyan Liao1, Kaiwen Zheng1, Xin Zeng1, Yuezhong Meng1, Hui Yan2, and Yang Liu3 1The Key Laboratory of Low-carbon Chemistry & Energy Conservation of Guangdong Province, and School of Materials Science and Engineering, Sun Yat-sen University, Guangzhou, China 2School of Computer Science, Sun Yat-sen University, Guangzhou, China 3School of Electronics and Information Technology, Sun Yat-sen University, Guangzhou, China hzh29@mail.sysu.edu.cn EDFAAO (2024) 2:10-18 1537-0755/$19.00 ©ASM International® INTRODUCTION The MicroStructural Hierarchy Descriptor (µSHD)[1–3] was initially proposed under the background of materials genome engineering (MGE),[4] where a quantitative description on material microstructure is critical to find the so called “materials genomes.”[5] To achieve this goal, quantitative relationships must be established between the processing, microstructure, and properties/performance of different materials. Therefore, the quantification of a microstructure, which to date is based only on crude low-level statistics, is an essential step, and µSHD has been proposed in this context. In the field of microelectronic failure analysis (MFA), characterization techniques such as scanning acoustic microscopy (SAM), magnetic field imaging (MFI), scanning electron microscopy (SEM), energy-dispersive x-ray spectroscopy (EDS), transmission electron microscopy (TEM), and others with spatial resolutions that span multiple scales, are routinely used to obtain spectra and images for fault localization and isolation. For diagnostics at the integrated circuit (IC) level, atomic-scale or even electron-scale characterization techniques are common. In contrast, it involves mainly mesoto-macroscopic scales at the packaging level. However, the bottleneck switches to packaging and interconnects in the 3D IC era. Recent years have seen the significant miniaturization of packaging and interconnects along with the IC processing node to enhance the performance and efficiency of electronic devices. Therefore, comprehensive characterization techniques that span from the electron scale to the macroscopic scale are required for MFA. In addition, the task of fault localization and isolation is mainly experience-based, and the data from characterization are only qualitatively utilized according to the operator’s experience and background. However, artificial intelligence (AI) may automatically control and perform characterization equipment and MFA in the foreseeable future.[6–8] With this vision in mind, a systematic and data-driven approach will be necessary. To this end, the convergence of the MGE and MFA objectives, and the µSHD approach will play a role in facilitating the adoption of AI in these two fields. It is noted that common AI tools have already been explored in the field of MFA for higher resolution imaging, increased throughput, enhanced image contrast, and faster scan times.[9] Deep convolutional neural network (DCNN)- based computing models have also been utilized for superior image segmentation and object classification.[10] While satisfactory results may have been obtained, the working principles behind DCNNs remain mysterious. It is an interpretative descriptor, i.e., the µSHD, rather than the mysterious one generated and hidden in the DCNNs, that is the focus of this work. The following sections introduce the concept of µSHD and benchmark its behavior for images captured with common characterization techniques to stimulate further and systematic application scenarios. MICROSTRUCTURAL HIERARCHY DESCRIPTOR As mentioned in the previous section, there are abundant spectra and images with resolutions of several orders
RkJQdWJsaXNoZXIy MTYyMzk3NQ==