SWIR Geiger-mode APD detectors and cameras for 3D imaging

Transcription

1 SWIR Geiger-mode APD detectors and cameras for 3D imaging Mark A. Itzler,* Mark Entwistle, Uppili Krishnamachari, Mark Owens, Xudong Jiang, Krystyna Slomkowski, and Sabbir Rangwala Princeton Lightwave Inc., 2555 US Route 13 South, Cranbury, NJ 8512 ABSTRACT The operation of avalanche photodiodes in Geiger mode by arming these detectors above their breakdown voltage provides high-performance single photon detection in a robust solid-state device platform. Moreover, these devices are ideally suited for integration into large format focal plane arrays enabling single photon imaging. We describe the design and performance of short-wave infrared 3D imaging cameras with focal plane arrays (FPAs) based on Geigermode avalanche photodiodes (GmAPDs) with single photon sensitivity for laser radar imaging applications. The FPA pixels incorporate InP/InGaAs(P) GmAPDs for the detection of single photons with high efficiency and low dark count rates. We present results and attributes of fully integrated camera sub-systems with 32 x 32 and 128 x 32 formats, which have 1 μm pitch and 5 μm pitch, respectively. We also address the sensitivity of the fundamental GmAPD detectors to radiation exposure, including recent results that correlate detector active region volume to sustainable radiation tolerance levels. 1. INTRODUCTION High-performance three-dimensional (3D) imaging technologies promise to deliver disruptive capabilities by providing a much greater wealth of information than that obtained from conventional two-dimensional (2D) imaging techniques. In particular, explicit high-resolution range (i.e., depth) information can eliminate ambiguities that are often present in 2D imagery, and 3D imaging enables advanced functionality such as the imaging of targets behind foreground obscurants that is generally not possible using traditional 2D intensity imaging approaches. Laser radar techniques, which employ optical signals with micron-scale wavelengths, allow for vastly higher resolution than traditional radar measurements. With sub-nanosecond scale laser pulse widths, range resolution on the scale of centimeters can be readily achieved. Furthermore, by using arrays of detectors with single-photon sensitivity, laser radar systems can operate at very large stand-off distances with very high area coverage rates. Focal plane arrays based on single-photonsensitive Geiger-mode avalanche photodiodes (GmAPDs) were first implemented in laser radar systems about a decade ago [1,2], and more recent system demonstrations have shown the ability to map hundreds [3] to thousands [4] of square kilometers per hour, which represents as much as an order of magnitude advantage in mapping rate over other competing laser radar systems. Initial pioneering work on GmAPD-based FPAs was carried out in the early 2s by researchers at MIT Lincoln Laboratory [5,6,7]. More recently, the industrialization of this technology by the present authors has led to the development of high performance sensors [8,9] with commercial availability of these sensors in turn-key camera systems [1]. Core camera functionality is provided by arrays of GmAPDs hybridized to CMOS read-out integrated circuits (ROICs) that provide independent time-of-flight measurements for each pixel and operate at frame rates as high as ~185 khz for 32 x 32 sensors and ~115 khz for 128 x 32 sensors. For both formats, average single-photon detection efficiencies beyond 4% (including all optical transmission and microlens array losses) are achieved for average dark count rates of less than 1 khz. Near-neighbor pixel optical crosstalk is <1%, and timing jitter is < 5 ps rms. Robust hermetic packaging of the FPA sensor provides high reliability operation in harsh environments, and the camera subsystems based on these FPAs incorporate FPGA-based sensor electronics, flexible GUI software, and industry-standard data interfaces. A key attribute on our roadmap for future generations of GmAPD 3D imagers is the push to larger array formats. Larger formats will necessarily be accompanied by reduced pixel pitch, which will in turn drive the implementation of smaller volume detectors in each pixel. Aside from enabling smaller pitch arrays, the reduction of the active volume of these detectors offers improvements in many pixel performance parameters that scale with active region size. Dark *mitzler@princetonlightwave.com; tel: ; Approved for Public Release, Distribution Unlimited.

2 count rate can be expected to reduce proportionally with the detector volume, and optical crosstalk between pixels which results from single photon emission during avalanche events also scales with the cross-sectional area of the active regions. Additionally, the radiation tolerance of these detectors is expected to scale with device active volume, and this issue is particularly critical in considering the deployment of GmAPD array-based sensors on space platforms supporting both imaging and free space communications applications. The primary challenge posed by implementing smaller active regions is greater difficulty in maintaining high optical coupling efficiency. In this paper, we describe the state-of-the-art in the performance of GmAPD-based 3D imaging cameras with 32 x 32 format as well as our more recent development of cameras with a 128 x 32 format. We present camera-level performance maps for dark count rate (DCR) and photon detection efficiency (PDE), as well as results for optical crosstalk. We then summarize the results of recent work done to fabricate GmAPDs with reduced active volumes and study the scaling of their performance for dark count rate and radiation tolerance. In particular, we find a close correlation of radiation tolerance levels with the inverse area of the detector active region. For GmAPD detector designs that are optically compatible with our existing FPA platforms, these discrete device results indicate that GmAPD sensors can be designed to tolerate radiation exposure of at least 8 krad with acceptable levels of dark count rate elevation. 2. GEIGER-MODE APD FPA AND CAMERA DESIGN 2.1 GmAPD Device Design and Integration into Photodiode Arrays The single-photon sensitivity of the sensor arrays described in this paper is provided by high-performance GmAPDs. The schematic cross-section at the left side of Figure 1 illustrates the basic design of the active device in each pixel. The wavelength range of sensitivity is determined by the material composition of the absorber region. We have developed cameras optimized for use with laser wavelengths between ~.9 and 1.1 μm in which an appropriate quaternary InGaAsP absorber is employed, as well as cameras with sensitivity between.9 and 1.6 μm compatible with more eye-safe sources near 1.5 μm in which a ternary InGaAs layer provides absorption over this wavelength range. A second critical region of the device structure is the multiplication region in which avalanche multiplication occurs. Operation of an APD in Geiger-mode entails biasing the device above its avalanche breakdown voltage into a metastable state in which a single photo-excited carrier can trigger a macroscopic pulse of avalanche charge that is GmAPD schematic device design Indium bump GmAPD pixel 32 x 32 GmAPD array anode contact SiN x passivation p + -InP diffused region i-inp cap multiplication region n-inp charge n-ingaasp grading i-ingaasp absorption n + -InP buffer n + -InP substrate anti-reflection coating cathode contact optical input Electric field Figure 1. Schematic illustration of GmAPD device design (left) incorporated into a single array pixel (center) which is ultimately integrated into a larger format photodetector array (right).

3 easily sensed using threshold detection electronics connected to the APD. Following the threshold detection, the avalanche must be quenched. In this mode of operation, strictly speaking there is no well-defined gain since the avalanche is self-sustaining and the total number of carriers generated during an avalanche depends more on the quenching circuit than on the properties of the device itself. However, we can estimate an effective gain of ~1 5 to 1 6 carriers generated during each avalanche prior to quenching. In order to establish the internal electric field profile required for Geiger-mode operation (e.g., as illustrated to the right of the device schematic in Figure 1), a suitable distribution of dopant charge in the APD structure is critical. This basic APD device structure is incorporated into an array geometry in which isolation trenches are etched between neighboring pixels (see center of Figure 1) to reduce optical crosstalk, which is discussed further below. Indium bumps are deposited on the anode contact of each active region to facilitate hybridization of the photodiode array (PDA) to an appropriate readout integrated circuit (ROIC) by flip chip assembly. Cathode connections are provided around the perimeter of the pixel array. A photograph of a completely processed 32 x 32 PDA is shown at right in Figure Design of GmAPD Focal Plane Array The principal elements of the assembled FPA module are illustrated schematically at the left side of Figure 2. The imaging functionality of the module is dictated by three semiconductor chips: (i) the InGaAsP GmAPD PDA just described; (ii) a custom silicon CMOS ROIC; and (iii) a GaP microlens array (MLA). Following the hybridization of the GmAPD PDA to the CMOS ROIC, we align and attach the MLA to the back surface of the PDA to increase the overall FPA optical fill factor. Based on the characterization of FPA pixel-level photon detection efficiency before and after MLA attachment, the effective fill factor for broad illumination of the FPA is found to be ~75%. Electrical connections to the ROIC bond pads are provided by automated wire bonding to corresponding signal traces on the ceramic interposer board (center of Figure 2). The interposer traces are then wire-bonded to a bond pad shelf in the ceramic housing, with connections leading to the pin grid array in the base of the package. A two-stage thermoelectric cooler maintains a temperature differential of up to 55 C; with a typical camera chassis ambient temperature near room temperature, the detector array in the FPA is generally operated at a temperature in the range of -25 C to -2 C. The fully assembled FPA is shown on the right side of Figure 2 just prior to attachment of the hermetic lid with a window allowing optical access to the sensor array. 2.3 Camera Construction and Functionality The FPA serves as the sensor engine of the camera head, which consists of a modular design based on three circuit boards: (i) the FPA board contains the FPA sensor and circuitry to control its power and temperature regulation; (ii) the FPGA board contains an Altera FPGA and a microcontroller that provide extensive on-board functionality through firmware programming; and (iii) the interface board provides an industry-standard CameraLink interface to the system computer, power regulation, and external clock and trigger inputs. The camera head requires only a single DC source between 12 and 36 V, from which all required biases and power levels are generated internally. The camera head is controlled by a high-performance personal computer with flexible GUI software, which includes algorithms and data structures to support the real-time storage of raw sensor data that reaches rates in excess of 4 Gb/s when the sensors are Lid MLA GmAPD PDA ROIC Ceramic interposer TEC Housing Figure 2. Schematic illustration of element stack-up in FPA assembly (left), with photograph of assembled chip stack on interposer (center) that is incorporated into full FPA sensor package (right).

4 running at their maximum frame rates. The raw data from the camera consists of 13-bit timestamps indicating photon time-of-arrival at each pixel to a resolution as high as.25 ns. Once a pixel fires during a given frame, it is disabled until all the pixels are re-armed for the next frame. Pixels that do not detect an event during a given frame will output the terminal count of the pixel-level counters. The maximum number of time bins during each frame is 8, and the highest resolution operation (i.e.,.25 ns time bins) corresponds to 2 μs range gates. A range extension function allows the camera to function with longer time bins so that the 32 x 32 camera can operate with range gates as long as 1 μs, and the 128 x 32 camera can have range gates up to 4 μs in duration. These range gate settings can be changed on a frame-by-frame basis X 32 GMAPD CAMERA PERFORMANCE x 32 Camera DCR and PDE Performance The most fundamental trade-off in GmAPD operation is that between the DCR and PDE. Higher PDE can be obtained by applying a larger excess bias voltage (i.e., the amount of voltage by which the bias exceeds the APD breakdown voltage), but this results in a higher DCR as well. The DCR performance of all pixels in a 32 x 32 camera operated at -25 C is illustrated in Figure 3 by an array-level DCR map along with the accompanying histogram. The average DCR Number of Pixels Avg DCR = 2.2 khz σ(dcr) =.4 khz Avg PDE = 3.5% Dark Count Rate (khz) Figure 3. Dark count rate performance for 32x32 GmAPD camera operating at 1.6 μm Number of Pixels Figure 4. Photon detection efficiency performance for 32x32 GmAPD camera operating at 1.6 μm % Avg PDE = 3.5% σ(pde) = 4.% Avg DCR =2.2 khz 2% 24% 28% 32% 36% 4% Photon Detection Efficiency 44%

5 of this array is 2.2 khz with an rms variation of.4 khz for an operating point corresponding to an average PDE of 3.5%. This average DCR corresponds to <.5% probability that any given pixel of the array registers a dark count during a 2 μs range gate. As the histogram shows, every pixel in the array has a DCR of less than 4 khz, and the tight distribution is roughly Gaussian. Residual performance non-uniformity is related to both epitaxial material variation, which is generally quite gradual and occurs over distances of many pixels, as well as process-related variation, which is responsible for the slightly lower DCR typically observed around the perimeter of the array. A corresponding performance map for the PDE of this camera is illustrated in Figure 4. PDE data were obtained using a pulsed diode laser source collimated across the entire FPA and calibrated to.1 photon per pixel area per pulse with a calibration uncertainty of ~5%. The rms variation about the measured average PDE of 3.5% is 4.%, and this variation is qualitatively similar to that seen in the DCR performance. These PDE values are asmeasured and account for all optical losses, including the less-than-unity fill factor of the microlens array. To demonstrate the evolution of the GmAPD FPA performance, in Figure 5 we summarize the DCR characteristics of 18 cameras plotted in chronological order of fabrication and representing several fabrication lots of GmAPD PDAs. The solid circles indicate the average DCR of each array, and the error bars indicate the standard deviation (±σ) of the DCR distribution for each array. The average DCR has tended to decrease appreciably, and the average ratio of the standard deviation to the mean is σ/dcr ~.3. In addition to inherent improvements in the DCR vs. PDE tradeoff at the GmAPD device level, improved optical coupling of the microlens array to the PDA allows a target PDE value (e.g., 3%) to be a achieved at a lower excess bias with a consequently lower DCR x 32 Camera Crosstalk Performance Dark Count Rate (khz) Figure 5. Dark count rate (DCR) data for 18 cameras corresponding to PDE = 31 ± 1% and operating temperature 248 ± 5 K. Solid circles ( ) indicate the average DCR over the array and error bars indicate the standard deviation (±σ) of the DCR distribution. During avalanche events in GmAPDs, charge acceleration in the multiplication region can give rise to hot carrier luminescence that results in the emission of photons. Although the probability of photon emission is low on the order of one photon per 1 4 to 1 5 charges the fact that all neighboring pixels are sensitive to single photons creates the possibility of crosstalk in which emitted photons trigger temporally and spatially correlated detections at neighboring pixels due to a primary pixel avalanche initiated by a signal photon. The coupling of crosstalk photons can occur by line-of-sight travel to nearest-neighbor pixels or by reflections from the back side of the PDA substrate to more distant neighbors. The spectrum of these emitted photons has been found to have a large peak at a wavelength corresponding to an energy just below the bandgap of the InP multiplication region material, as well as a longer wavelength blackbody tail [11]. We have employed two approaches to measuring the spatial dependence of crosstalk. One technique involves illumination focused to a single pixel, with a short pulse incident on this single pixel at a specific time within each successive frame. For those frames in which the illuminated pixel fires at the time of the pulse arrival, we then analyze the timing data for all other pixels to see if any of them fired a short time thereafter. The second technique for extracting crosstalk behavior is based on just the consideration of dark count data: instead of creating primary avalanches initiated by illumination, one can consider dark counts to be primary avalanches and then look for correlated counts at short time intervals later, just as in the illuminated case. In both techniques, the analysis of the resulting data depends on the choice of time frame considered to be temporally correlated, as well as the choice of distance considered to be spatially correlated. In previous preliminary work on crosstalk behavior, we have confirmed that both techniques yield comparable results [12]; however, the dark count analysis is more convenient because it does not require optical coupling to a single pixel, and it allows us to use events from every single pixel in the array Sensor Number Average of 32x32 pixels Errors bars indicate ± σ PDE = 31% ±1% T= 248K ±5K

6 In Figure 6 we present the results of crosstalk analysis for 1, frames of dark count data from a 32 x 32 camera. The analysis consists of identifying any dark counts that occur within 1 ns of each other. The map at the left of the figure is a compilation of all data in which we have assumed that the primary avalanche occurs in the center of this map (i.e., at the blue pixel labelled with ). The values and corresponding color coding indicate the relative incidence of crosstalk events at any given distance and direction from the primary avalanche. There is a geometric pattern to the data that is found to be very consistent and is related to specific geometric features in the arrays themselves, e.g., attributes of the etched isolation trenches and the nature of reflections from different positions on the back surface of the PDA. Although we have reduced line-of-sight coupling of photons to adjacent pixels by etching optical isolation trenches between them, the crosstalk between these near-neighbor pixels is still dominant. Trenches are etched through the entire epitaxially grown structure (i.e., through the absorber), and the anisotropy in this near-neighbor coupling is related to the crystallographic dependence of the InP wet-etch chemistry used, resulting in a V-groove geometry in one direction and a nearly vertical dove-tail geometry in the orthogonal direction, as described previously [14]. A B Figure 6. Crosstalk performance for 32x32 GmAPD camera operating at 1.6 μm at 3% PDE. At left is a spatial map of crosstalk events from 1, frames of data. Values represent number of crosstalk counts at each location within a 15 x 15 pixel area. The primary avalanche pixel is represented by the blue square at the center. At right is an llustration of crosstalk photon propagation explaining structure in spatial map. From a primary avalanche at pixel 1, line-of-sight coupling from 1 to 2 is reduced by etched isolation trenches. Reflective coupling from 1 to 4 is reduced by an absorptive metallic coating at A. The requirement for a dielectric aperture at B to allow signal photons (incident from top of figure) to reach pixel 3 promotes relatively high reflection from 1 to 5. Beyond the eight line-of-sight near-neighbors, crosstalk coupling to further neighbors occurs by reflection from the back surface of the PDA substrate, as shown schematically at right in Figure 6. Such reflections have been reduced by depositing an absorptive metallic film over as much of the PDA back surface as possible. However, it is necessary to have dielectric anti-reflection coatings (ARCs) within apertures placed directly above each pixel to facilitate coupling of light collected from each microlens in the MLA to its corresponding PDA pixel active region. For large angles of incidence, these ARCs give rise to high reflectance and define symmetry points for significant reflections to further neighbor pixels. These reflection characteristics are illustrated schematically at right in Figure 6 and explain the highlevel symmetry seen in the spatial crosstalk map at left in the figure. Total cumulative crosstalk within a 1 ns time window and the illustrated 15 x 15 pixel region is less than 8%.

7 X 32 GEIGER-MODE APD CAMERA PERFORMANCE x 32 Camera DCR and PDE Performance Characterization of the DCR and PDE performance for the 128 x 32 format cameras is analogous to that described in the previous section for the 32 x 32 cameras. In Figure 7, we show the array-level maps for both DCR and PDE, along with corresponding histograms of their performance distribution, for 1.6 μm arrays operated at -2 C. Given that the 5 μm pixel pitch for these arrays is half that of the 32 x 32 arrays, the direction in the array with four times as many pixels has a physical dimension twice as large (i.e., 6.4 mm) as the dimensions of the smaller arrays (i.e., 3.2 mm). Along the length of this longer dimension, the array is more susceptible to pixel performance variation. However, as is evident in the performance maps in the figure, we are still able to obtain very good uniformity of both DCR and PDE in this dimension. The histograms exhibit reasonably Gaussian distributions, and for an average PDE of 33.4%, we obtain an average DCR of 6.8 khz. The 18 μm diameter active regions of the 128 x 32 FPAs have an inherently lower DCR than the 34 μm diameter active Figure 7. Dark count rate (top) and photon detection efficiency (middle) performance maps, along with corresponding performance histograms (bottom) for 128 x 32 GmAPD camera operating at 1.6 μm and -2 C ## ## ## ## ## ## ## ## ## ## ## ## ## ## ## ## ## ## ## ## ## ## ## ## ## ## ## ## ## ### ### ### ### ### ### ### ### ### ### ### ### ### ### ### ### ### ### ### ### ### ### ### ### ### ### ### ### ### ## More Number of Pixels Dark Count Rate (khz) Avg DCR = σ(dcr) = Avg PDE = 6.8 khz 3. khz 33.4% % 19% 22% 25% 28% 31% 34% 37% 4% 43% 46% 49% 52% Number of Pixels Photon Detection Efficiency Avg PDE = σ(pde) = Avg DCR = 33.4% 5.% 6.8 khz

8 regions of the 32 x 32 FPAs. However, much tighter optical tolerances and consequently higher microlens array coupling losses for the smaller detectors in the 128 x 32 FPAs require operation at a higher excess bias voltage to achieve the same PDE of ~3 35%. This is the dominant factor explaining the higher average DCR value found in Figure 7 for the 128 x 32 FPA relative to the DCR seen in Figure 3 for the 32 x 32 FPA x 32 Camera Crosstalk Performance The crosstalk performance of the 128 x 32 FPA is qualitatively similar to that demonstrated for the 32 x 32 FPA in the previous section, as is seen by comparing the data presented in Figure 8 with that in Figure 6. However, if device designs were equivalent, the crosstalk of the 128 x 32 array would be higher because of the closer spacing of its active regions (as dictated by halving the pixel pitch). Although the fact that the active regions of the 128 x 32 FPA have a smaller diameter 18 μm, as opposed to the 34 μm diameter of the 32 x 32 FPA active regions would tend to decrease the crosstalk approximately linearly with the diameter for line-of-sight coupling, the solid angle that the active region presents to emitted crosstalk photons scales as 1/R 2 where R is the distance between pixels. This explains a factor of two increase in the nearest neighbor line-of-sight crosstalk probability per pixel. However, the crosstalk at further neighbors has been effectively reduced by incorporating a suitable epitaxial filter layer [11] between the substrate and buffer of the GmAPD structure (refer to the schematic at the left side of Figure 1) so that a significant fraction of the crosstalk photons are absorbed before they can be reflected, without attenuating the input signal photons. The cumulative crosstalk within at 15x15 pixel area is under 1% for an average array PDE as high as 3% Figure 8. Crosstalk performance for 128x32 GmAPD camera operating at 1.6 μm. See caption of Figure 6 for explanation. Total cumulative crosstalk within a 1 ns time window and a 15 x 15 pixel region is less than 1%. 5. REDUCED VOLUME GMAPD DETECTORS AND RADIATION TOLERANCE A key element of any general roadmap for the development of future generations of GmAPD 3D imaging sensors will include larger format arrays. Scaling to larger formats will require further reduction in pixel pitch and a corresponding decrease in the size of the detector active region in each pixel. Even for arrays of the current generation formats (e.g., 32 x 32 and 128 x 32), there is additional motivation for the implementation of small active regions because many performance parameters improve with the reduction of device size. In particular, dark count rate can be expected to scale with the volume of the active region. Optical crosstalk between pixels should also reduce because the cross section presented by neighboring pixels will shrink. A similar argument applies with regard to radiation tolerance since a smaller device volume will have a lower probability for experiencing radiation damage given a fixed particle fluence. The primary tradeoff encountered with smaller active regions is the greater difficulty in achieving high efficiency optical coupling. In the context of GmAPD arrays, this poses a challenge in the design, alignment, and attachment of microlens arrays consistent with high effective fill factor in focal plane arrays with smaller active regions in each pixel. 5.1 DCR vs. PDE Improvement with Reduced Volume We have developed a number of approaches for fabricating GmAPDs with reduced active volumes, and while there are details beyond the scope of this paper, one general approach is to employ the fundamental device design summarized in