1. Introduction

Precise timing management of short optical pulses is indispensable in a wide range of applications, such as ultra-high-speed optical networks [1], optical signal processing [2], optical measurement [3], and so forth. In addition, real-time timing measurement is important for applications such as system stabilization [4], synchronization [5], and optical channel monitoring. To handle such high-speed phenomena while satisfying the application requirements, two-dimensional (2D) non-pixilated optical switches, such as organic-dye molecular thin films, are one emerging enabling technology capable of ultrafast optical response [6, 7]. Using the spatial parallelism of such films while satisfying the precision and real-time timing measurement requirements, among others, we have reported a timing measurement technique for jitter and skew reduction (namely, deskewing) [8] using a surface-normal, two-dimensional (2D) ultrafast all-optical switch formed of squarylium dye J aggregate (SQ-J) [6]. With this technique, the timing of short optical pulses was converted to space-domain position, which was physically detectable with a sensor array. Moreover, it was possible to simultaneously determine the timing of multiple optical pulses owing to the large-area, two-dimensional structure.

However, as described in detail below, the bandwidth of the system was in fact limited by the data acquisition and serial computation procedure even though the 2D all-optical switch itself had extremely fast response and massive parallelism. In this paper, considering that the space-domain processing could be inherently performed in a parallel manner, we introduce the idea of on-chip parallel processing to such a short optical pulse measurement system. This provides a smooth flow of information from the ultrafast phenomena, which is undetectable without ultrafast all-optical switches, to higher-layer silicon-based complex serial signal processing, via an optoelectronic parallel computational sensor array.

2. Parallel architecture for short pulse measurement

We first describe the principle of ultrashort pulse detection. The left-hand side of Fig. 1, denoted by “all-optical”, extracts the timing of short pulses using a 2D optical switch based on a simple time-to-space conversion [3, 9, 10]. Here, the complete wavefront of a horizontally spread-out signal pulse is incident on the film simultaneously. A control pulse, on the other hand, is obliquely incident upon the film. The control pulse, when intersecting the film, makes a transparent region that effectively works as a slit allowing the signal pulse to pass through. Therefore, the arrival timing of the signal pulse can be estimated by determining the position at which it is transmitted. Owing to the ultrafast optical response (around 100 fs) and large area of the film, it is possible to detect the timing of multiple pulses incident at different locations on the film.

 

Fig. 1. Optical pulse timing measurement by an ultrafast all-optical 2D switch and an optoelectronic computational sensor array.

Download Full Size | PDF

The issue this paper address is the right-hand side of Fig. 1, indicated by “optoelectronic”. If we use a conventional CCD camera for detection of the output optical pulses, the frame-rate (typically 30 frames-per-second (fps)) severely limits the latency and throughput of detectable timing information. A non-scanning photodetector array would be a better choice, but still, there would be associated bottlenecks, such as wiring complexity, in order to realize high-density, massively parallel, two-dimensional operation [11, 12]. It is required for both cases, i.e. single and multiple signal pulses, to compute the pulse timing using an array of detectors, which involves a certain computation latency. Notice the fact that the timing extraction procedure in this architecture essentially has spatial parallelism: timing extraction is equivalent to position estimation of spots in the space-domain; the processing can be carried out entirely in parallel. Therefore, to achieve high-density and high-throughput detection of optical pulse timing in the space-domain, one idea is to integrate the computation capability in parallel in the space-domain; to this end, we adopted a parallel computational sensor array where photodetector and computational circuits are integrated [11]. If all data captured by a conventional camera were serially transmitted to a higher-level processor, unavoidable longer latency and reduced throughput would result. On the other hand, a computational sensor array is a device in which photo-detection and a certain computational capability are integrated at each pixel. Hence, the data flowing from the sensor to higher-level systems includes only the extracted timing information and, therefore, the volume of data is much less, which reduces the overall latency and enhances the throughput of the timing extraction. Application of so-called smart pixels to ultrafast optics was mentioned in [13], and silicon VLSI circuits have been coupled to high-speed, all-optical demultiplexing switches in, for instance, optical label recognition applications [14]. Compared to those studies, this paper, toward building a hierarchical system, demonstrates a parallel algorithm based on ultrashort pulses using high-density nearest-neighbor-connected parallel processing resources, as discussed in detail in the next section.

3. Algorithms and experimental demonstrations

To verify the principle, we performed experiments using pulses generated by a Ti:Sapphire regenerative amplifier system with an optical parametric amplifier. The wavelengths of the signal pulse and control pulse were 760 nm and 800 nm, respectively. The pulse width was 100 fs and the repetition rate was 1 kHz. The energy density of the control pulse was around 4 µJ/mm², which induced a transmission efficiency on the order of 10^-5 in the SQ-J film. Timing fluctuations were applied to the input signal pulses by changing the time-of-flight using a retroreflector attached to a piezoelectric actuator. We introduced a computational sensor array, manufactured by Hamamatsu Photonics K. K. [15]; this device includes a computation circuit and an integrated optoelectronic chip composed of 128×128 photodetectors and 1×128 analog-to-digital converters. It essentially has an architecture where each pixel has its own photodetector and computation capability. In addition, each pixel is connected to its four neighboring pixels. Therefore, two-dimensional parallel processing of data is possible. In the particular device we used, the computation function was implemented in a field-programmable gate-array (FPGA) for cost efficiency. The device had a 1000-fps bandwidth. Since the experimental system used here had a 1-kHz pulse repetition rate, the device could estimate the timing of the incoming signal pulses on a pulse-by-pulse basis. It should be noted that the system can detect timing of pulses whose transmission rate is in the terabit per second range as long as their timing fluctuation dynamics are within the limitations of the sampling rate (theoretically 500 Hz), since the film has the capability to handle 1-ps interval pulses, as demonstrated in [16]. The pixel pitch of the photodetector array was 40 µm, and the aperture ratio of each pixel was 70%. The photodetector array was coupled to a micro-channel plate, which had a luminance gain of 12,000 lm/m²/lx, via a fiber optic plate [17].

We could operate the computational sensor array in two different modes: One is a “dumping mode” where all of the sampled data is transferred directly to the host computer without performing any processing. The other is a “computation mode” where only the results of computation are transferred to the host computer. Using the on-chip computation circuits, it is capable of calculating feature values, such as the first-order momentum, which provides the center of gravity of an image with sub-pixel precision.

The calculations performed by the system are described as follows. Suppose that N distinct optical pulse patterns are obtained by the photodetector array. The device then finds pixels that give local intensity maxima, denoted d by (x_i, y_i) (i=1, …N), by communicating with neighboring pixels. Then, for each of those positions, the following two values are computed within the device: ∑_Mw=-M^s(x_i+w,y_i) and ∑_Mw=-w×s(x_i+w,y_i), where s(x_j,y_j) represents the pixel value at (x_j,y_j). Note that these computations are performed in parallel for N spots. The values are then transferred to a higher-level processor, and the center of gravity along the horizontal direction is calculated as

The pulse timing is then derived as

where c is the speed of light and θ is the intersection angle of the control pulse and the film. M was equal to 2 for the device used. The device can extract a maximum of 512 different spots, which is determined by the available memory on the chip. The solid line in Fig. 2(a) shows experimental results of the calculated center of gravity, in other words, the estimated timing, as a function of timing difference given by changing the arrival timing of the signal pulse. The dashed line in Fig. 2(a) represents the standard deviation of the estimated position, which is about 0.1 pixel. Therefore, if the system can resolve a 0.2-pixel difference in the center of gravity, it can resolve a timing difference of about 10 fs since one pixel corresponds to about 50 fs in this example.

 

Fig. 2. (a) Experimental results of the relationship between timing difference and extracted pulse position (center of gravity) in the space domain. (b) A simulated space-domain profile of an optical pulse. (c) Peak and center of gravity positions with different input-signal pulse widths.

Download Full Size | PDF

In addition to the fact that calculating the center of gravity from the output pulses allows sub-pixel precision, as described above, it also endows the timing estimation with robustness to variance of the input pulse width. This is explained as follows. We denote the transmission dynamics of the SQ-J film, in other words, the impulse response of the system, as f(t). Let the temporal intensity profile of a control pulse be p(t). The transmission of the SQ-J film at the horizontal position x is then given by a convolution of the film’s impulse response and the control pulse intensity, as

Since the input signal pulse can pass through the film depending on the transmission of the film, the time-space distribution of the transmitted optical intensity is given by a multiplication of (3) and the profile of the input signal pulse. Here we simulate f(t) as an exponential decay given by f(t)exp(-ln2(t/t_o)), where t ₀ is the response time of the film. We estimated the value of t ₀ based on the curve in Fig. 4 of Ref. [6], by assuming that the curve is a convolution of f(t) and a Gaussian input pulse. The half-maximum decay time agrees well with the case when Δt ₀=74 fs. Then, we can estimate the spatial extent of the output signal pulse in the space-domain. We simulated the spatial profile and the position of the peak for different signal-pulse widths while keeping the control-pulse width constant (100 fs FWHM). The intersection angle between signal and control pulses was set to 13°. The solid curve in Fig. 2(b) shows a simulated space-domain profile when the signal pulse width is 100 fs. The position of its intensity maximum and that of its center of gravity are shown. As shown by the dashed line in Fig. 2(c), the peak-position differs with respect to the input signal pulse width. The center of gravity for each pulse, on the other hand, which is given by ∫xs(x)dx/∫s(s)dx, remains constant regardless of the signal pulse width, as shown by the dotted line in Fig. 2(c). The tendency was the same when the control pulse width was changed.

 

Fig. 3. (a) Consecutive snapshots of the pulses transmitted through the SQ-J film captured by a CCD camera (left, 33ms interval) and a computational sensor array in dumping mode (right, 1ms interval). (b) A 15-Hz, 600-fs timing fluctuation extracted by the computational sensor array in computation mode. (c) Movie captured by i) conventional CCD camera, and ii) computational sensor array (in dumping mode) [Media 1].

Download Full Size | PDF

For high-bandwidth extraction, Fig. 3(a) compares the data acquisition capability by snapshots of the output pulses transmitted through the film, where the horizontal position of each pulse corresponds to the timing difference. Here, the input signal pulses were made to experience a timing fluctuation of 15 Hz. The left-hand side of Fig. 3(a) shows the snapshots captured by a CCD camera (30 fps); note that we observed blurred images and the pulse positions in two successive frames were spatially shifted. On the other hand, in the right-hand side of Fig. 3(a), which was obtained by operating the computational sensor array in dumping mode, we observed clear pulses in successive frames. Using the computation mode of the sensor, we also obtained a trace of the timing, as shown in Fig. 3(b), where a 15-Hz, 600-fs timing fluctuation can be clearly observed. Figure 3(c) compares the performance in a movie consisting of two parts: i) Movie captured by a conventional CCD camera, and ii) Slow motion replay of the pulses obtained in dumping mode of the sensor. 15-Hz timing fluctuation is clearly observed when the computational sensor array is used.

High-bandwidth extraction of pulse timing allows efficient realization of a hierarchical system in which a higher-level processor takes the role of calculating secondary information based on timing information that has already been extracted by lower-level sensors. Figure 4 shows the results obtained by such a system, where the timing fluctuation contains two frequency components (5 Hz and 15 Hz). The pulses were extracted by the film, their timing was obtained by the computational sensor (Fig. 4(a)), and a higher-level system, in this case a microprocessor in the host computer, performed a Fast Fourier transform (FFT) to analyze the frequency components of the fluctuations. Two frequency components were successfully extracted, as shown in Fig. 4(b).

 

Fig. 4. Demonstration of hierarchical processing. (a) The pulse timing was obtained by the computational sensor array, and (c) FFT analysis was performed in the higher-level processor.

Download Full Size | PDF

4. Summary

In summary, a parallel architecture is proposed to obtain high-bandwidth timing measurement using parallel optoelectronic processing between ultrafast all-optical switches and higher-layer signal processing. Timing fluctuations of optical pulses, which are extracted by a 2D ultrafast all-optical switch and which are not normally detectable by conventional cameras, are captured and extracted using on-chip computation in the computational sensor array. The high-bandwidth timing measurement capability will enhance various timing management systems, such as deskewing systems [8]. Such a parallel processing approach would also be applicable to and significantly improve the dynamic performance of systems where time-off-light of short pulses is converted to space-domain position information, such as three-dimensional object height measurement [18]. The 2D optical switch used in this paper, the SQ-J film, is operated at a wavelength of about 800 nm, which is not well-matched with telecommunication applications, but another molecule has already been developed for operation at 1.55 µm [19, 20].