Aug_EDFA_Digital

edfas.org ELECTRONIC DEV ICE FA I LURE ANALYSIS | VOLUME 23 NO . 3 30 in standard time-integrated scheme (Fig. 4). Based on the device floorplan available in the manufacturer develop- ment tools, this area is related to the IO. The AOI was scannedwith a resolution of 32 x 32 pixels and a dwell time of 2 seconds per pixel. A total of 72,729 photonswere collected. Thephoton countinghistogramof the whole area shows five groups of spikes (Fig. 5). These match with the input signal of 5 MHz when the frequency of the reference signal is 1 MHz. A photon emission image (i.e., 2D photon histogram) was generated by counting the number of photons detected at a specific location/ pixel (Fig. 6). The resolution of this image is the same as the scanning resolution (32 x 32 pixels). On the figure, the coordinates are computed according to the field of view. STDBSCANclusteringwas appliedwith the parameters ( ε XY = 1 pixel, ε T = 15 ps and MinPts = 3). These parameters were chosen after observing and analyzing the distance to the nearest neighbors. After clustering, only 847 photons were considered as signals and classified into 216 clusters. The others have been labelled as noise and discarded. The clustering parameters to identify core points are quite stringent and as a result, many clusters were constituted of the minimum point requirements. It is possible that due to the multiple photon densities existing in this area, the user-defined parameters may have unintentionally discarded photons emitted from sources where emission was sparser in time or space. The pdfs were then computed for each cluster with an improved resolution factor of two and summed to generate a higher resolution image (Fig. 7). The image resolution is now 64 x 64 pixels instead of 32 x 32 pixels. Compared to Fig. 4(b) and Fig. 6, the emission from the spot around (20,10) in Fig. 7 seems less intense than the one from (30,20). Study of TRE waveforms in (20,10) prior to the process showed longer emission than in (30,20), meaning photons are more spread along the time axis and thus less dense in the ( x,y,t ) space. The chosen clustering parameters gave the advantage to the generation of clusters with little time spread and thus discarded many photons from the area located around (20,10). This issue could potentially be tackledbyusinga clusteringalgorithmcapableof handling multiple densities. Notwithstanding this matter, the spot located at (30,20) is nowbetter resolved andmore visible. This result shows that machine learning based process can help in identifying fainter emission spots in addition to generating higher resolution image fromdata acquired at lower resolution. This demonstrates the potential of the suggested approach. Fig. 5 Photon counting histogram (i.e., TRE waveform) for the whole area squared in red in Fig. 4(a). Histogram bin of 1 ns. Fig. 6 2D histogram generated from the 3D data with an image resolution of 32 x 32 pixels. Data acquiredwith the 50X objective lens equipped (NA 0.4). The color bar indicates the number of photons. Fig.7 Interpolated image from the results of clustering by STDBSCAN. Clustering parameters: EpsT = 15 ps, EpsXY = 1 pixel, Minpts = 3. Image resolution was interpolatedby a factor of two (new image resolution is 64 x 64 pixels).