May_EDFA_Digital

edfas.org ELECTRONIC DEVICE FAILURE ANALYSIS | VOLUME 22 NO. 2 34 energy at the frequencies 7 Hz, 24 Hz, 40 Hz, 58 Hz, and 81 Hz. The results show excellent agreement confirming the theoretical considerations above. Figure 8 shows an example of a 3Dhot-spot localization ina fullypackaged stackeddevice. Thedevice contains two dies with die attach in between, glued onto a lead frame andencapsulatedbymoldingcompound. Resistivedefects are present at both die levels. A schematic of the sample and the axial locations of the hot spots are provided in Fig. 8a. A lock-in measurement including the acquisition of the TRTR was performed at the frequencies of 1 Hz and of 4.22 Hz. Figure 8b contains a series of LIT amplitude images of the upper-die defect extracted from the time signal at increasing frequencies ranging from 1 Hz to 15 Hz. These parametric images were computed for up to the 14 th harmonic from the TRTR data recorded from a single LIT measurement at 1 Hz lock-in frequency. It should be noted, that the lateral imaging resolution of the hot-spot increases with increasing frequency. The corresponding phase-shift-to-frequency characteristics calculated from the 1 Hz and 4.22 Hz TRTRmeasurements are shown in Fig. 8c. The phase estimates fromthe two hot spots can clearly be separated. Also, the TRTR obtained from the deeper hot spot shows larger phase values as the delay is larger according to the longer distance the thermal signal needs to travel until reaching the sample surface. It can also be noticed that the standarddeviations of the phase values increasewith frequency and are larger for the hot spot that is located deeper inside the sample (lower die). The larger standarddeviation represents larger relative noise amplitudes which may have their origin in either a lower power consumption of the lower defect or higher attenuation the thermal signal was exposed to on its propagation path. Besides the advancements in extraction of the phase transfer function from a single LIT-measurement, the availability of the TRTR also opens paths for further signal analysis approaches. While conventional analysis is based on filtering or decomposing transformations, novel approaches based onmachine learning algorithms may also be applicable. Currently, ongoing research investigates the application of principal component analysis (PCA) for the separation of laterally overlap- ping thermal sources based on characteristic features of their temporally resolved thermal response. Results of this analysis approach are shown in Fig. 9. Figure 9a shows the amplitudes at the fundamental frequency of the TRTR data. In Fig. 9b, the weighting coefficients of a principal component extracted by PCA of the same data set is plotted. The presence of two hot spots (denoted by the arrows) can be identified in Fig. 9b, instead of a single larger one shown in Fig. 9a. SUMMARY AND CONCLUSIONS With the introduction of 3D architectures into micro- electronics components and systems, precise defect localization is requested to extend from 2D to 3D in order tomaintain theavailabilityof comprehensive failureanaly- sis for investigating and understanding the underlying defect-causes. Lock-in thermography (LIT) with its unique contrast mechanism is one of the few non-destructively operating techniques that can provide sufficiently high lateral and axial resolution for 3D measurements. This article describes an approach for 3D-localization of thermally active defects by lock-in thermography. It is based on the technique's superior sensitivity of LIT signals to temporal variationswith respect to the electrical stimu- lation. In this approach, the phase shift, which is related to Fig. 9 LIT analysis using machine learning algorithms: (a) amplitude image at the fundamental frequency of TRTRdata and (b) parametric image of theweighting coefficients of a principal component of the TRTR- data set. Two separated thermal sources can be identified (denoted by the arrows). (a) (b)