August_EDFA_Digital

edfas.org ELECTRONIC DEV ICE FA I LURE ANALYSIS | VOLUME 24 NO . 3 36 Fig. 8 Benchmarking between conventional XRM and the novel 3Dx-ray system. Imagingwith thenovel system showed that cracks in the traceandvoid in theorganic substrate are clearly identified, without noise or beam hardening artifacts. The acquisition time to observe these failures is only a fraction of that of the conventional XRM. to see this layer clearly using x-ray techniques, much less to be able to locate defects in this layer. The results of scans done at 0.3 μm resolution are shown in Fig. 5. Because of the low contrast of this layer, exposure time was increased so the high-resolution scan was arbitrarily set for 2.5 hr. In general, this acquisition time can be much shorter. A comparisonbetween exposure times of 6 and 87min. is illustrated with a virtual delayered slice from a series of solder joints in a cold-joint/non-wet detection scan as identified by the meniscus like curvature contact. The 6min. scan is sufficient for failure detection, without com- promising much loss of image clarity (Fig. 6). ELIMINATION OF BEAM-HARDENING ARTIFACTS One of the key advantages of this tool is the acquisi- tion of clean and clear images that are free from beam hardening artifacts such as streak lines and bands. To illustrate this, the team ran the same sample showing slices at the under-bump metallization (UBM) and RDL interface, scanned with the Apex and with a con- ventional XRM, model PrismaXRM as a benchmark (Fig. 7). The results showed a remarkable suppression of beam hardening artifacts by the novel x-ray system. Beam hardening artifacts and noise are particularly detrimental when trying to identify small defects in the vicinity of those streak lines or bands. This is illustrated in the following comparison between the novel 3D x-ray system and the conventional XRM (Fig. 8). TRACE CRACKS, VOIDS IN ORGANIC SUBSTRATE While some of the above data were scanned at 30 min. to 2.5 hr, in principle, most of the defects can be identified within a few minutes. Time to identify a defect quickly is most critical in a pos- sible application to evaluate 300 mm wafer-level defects duringmanufacturing where a defect can be identified within a few minutes. A proof of concept to evaluate a 300mm wafer for voids in through silicon vias (TSV) is shown in Fig. 9. With the imple- mentation of machine learning and AI, the group believes this technique can be deployed in mass production for wafer level inspection. Fig. 9 3D data set showing local variations in metallization density at the edge of a 300-mmwafer. Dataset was collected in ~15 minutes with a voxel size of 0.5 µm.