Aug_EDFA_Digital

edfas.org ELECTRONIC DEV ICE FA I LURE ANALYSIS | VOLUME 23 NO . 3 24 MACHINE LEARNING FOR TIME-RESOLVED EMISSION: IMAGE RESOLUTION ENHANCEMENT Samuel Chef, Chung Tah Chua, and Chee Lip Gan Temasek Laboratories@NTU, Nanyang Technological University, Singapore csamuel@ntu.edu.sg EDFAAO (2021) 3:24-31 1537-0755/$19.00 ©ASM International ® INTRODUCTION Despite the challenges posed by continuous scaling of transistors and decrease in operating voltage, photon emission [1] remains one of the go-to techniques for defect localization in ICs. [2] For instance, recent work on resistive randomaccessmemory (ReRAM) showed its relevance on emerging technologies. [3] In parallel to failure analysis, there has been a growing interest from the hardware security community for this technique and its capability to extract encryptio n keys from ICs has been reported by several research gro ups. [4] Hence, photon emission is an important tool for IC analysis with many applications. Over the years, many different detection schemes have been developed. They can be differentiated between the time-integrated (such as camera array sensor) and the time-resolved schemes. [5] The former category can solely generate one type of outputs, an image that shows the location of emission spots. The second category, the time-resolved schemes, can be further divided between the time-resolved 2D array sensors [6] and the single point sensors coupled with a scanning head. [7] Time-resolved systems can provide additional information but they are more complex to set up compared to time-integrated sensors; they require an additional synchronization signal. However, studies such as those by P. Perdu et al. [6] and F. Stellari et al. [7] have demonstrated the benefits of time-resolved detection schemes for IC analysis and dis- cussed applications for the most advanced VLSI devices and circuits. As transistor scaling progresses, precise identification of locations of interest for defect localization on the latest technology nodes has been challengedby resolution limits of optical systems. Solutions have been suggested by the community to address this problem and they may be divided between the hardware and software approaches. The development of state-of-the-art solid immersion lenses with very high numerical aperture is an example of a hardware approach that greatly enhances the system capabilities. In addition to the hardware approach, soft- ware (also sometimes referred as post-acquisition pro- cessing) has also been explored by researchers to further extend capabilities of the technique. For instance, F. Lan et al. [8] reported the use of deconvolution algorithms to recover better resolved images. Another approach pre- sented by I. Vogt et al. [9] suggested to recover a higher resolution image from multiple im ages acquired from different viewing angles, like com puted tomography. F. Stellari et al. [10] reported an improvement of the local- ization of emission spots by computing the difference between emission images acquired at different time intervals of the test sequence and under various bias con- ditions. All these approaches mostly focus on processing signals in space dimensions (or images) and take little consideration of the time dimension. Another issue is that when working with commercial devices, the end-user may have limited control over the biasing conditions of the device under test (DUT). For instance, only the device I/Os are accessible. This does not allow the end-user to switch individual gates of the circuit and prevents some of the approaches discussed above. As a result, correlating emission spots with specific nodes of interest can be challenging if only considering the 2D signals/images. Finally, when using a scanner-based, single-point sensor scheme such as shown in F. Stellari et al., [7] scan- ning a complete area can take quite some time in order to ensure a large enough signal-to-noise ratio. A trade-off has to be found between short enough measurement time and sufficient signal-to-noise ratio. Typically, the measurement time is reduced by lowering the scan- ning resolution, i.e., reducing the number of pixels or