Preprint of SPIE paper 7706-2, to be presented at the Defense, Security, and Sensing Symposium, Orlando, FL, Apr. 8, 2010.

Design of a Wireless Sensor Network with Nanosecond Time Resolution for Mapping of High-Energy Cosmic Ray Shower Events Michael P. Frank*a, Sachin S. Junnarkarb, Triesha Fagana, Ray H. O’Neal, Jr.a, Helio Takaib a Dept. of Physics, Florida A&M Univ., Tallahassee, FL, USA 32307-3102; b Brookhaven National Laboratory, Upton, NY, USA 11973 ABSTRACT We describe a low-cost, low-power wireless sensor network we are developing for high time-resolution (ns-scale) characterization of particle showers produced by ultra-high-energy (UHE) cosmic rays, to infer shower direction at sites where hard-wired data connections may be inconvenient to install. The front-end particle detector is a scintillator block monitored by a photomultiplier tube (PMT). We keep the sensor nodes synchronized to within 1 ns using periodic highintensity optical pulses from a light-emitting-diode (LED) overdriven at very high current (~30 A) in short (4 ns) bursts. With minimal optics, this signal is resolvable under free-space transmission in ambient light conditions at multi-meter distances using a high-speed avalanche photodiode (APD) receiver at each node. PMT pulse waveforms are digitized relative to this precise time reference on a Field Programmable Gate Array (FPGA) using a Time-over-Threshold (ToT)/Time-to-Digital Converter (TDC) digitizer developed at BNL. A central server receives timestamped, digitized PMT pulse waveforms from the sensor nodes via Wi-Fi and performs real-time data visualization & analysis. Total cost per sensor node is a few thousand dollars, with total power consumption per sensor node under 1 Watt, suitable for, e.g., solar-powered installations at remote field locations. Keywords: wireless sensors, sensor networks, Wi-Fi, radiation detection, particle detection, scintillator detector, LEDs, optical synchronization

1. INTRODUCTION In this section, we present some of the general background and motivation for this work. 1.1 Open Problems in Cosmic Ray Astronomy The origin of the ultra high energy cosmic rays (UHECR) remains a mystery which continues to challenge astrophysics[1],[2],[3],[6],[18]. These particles represent energies beyond the Greisen-Zatsepin-Kuz'min (GZK) cutoff which is the energy beyond which interaction with the cosmic microwave background (CMB) radiation dissipates the energy of these particles from pion photoproduction[9],[21]. Upon entering the atmosphere of the Earth, these high energy particles, consisting primarily of protons or gamma rays (high energy photons), initiate a shower of secondary particles from interaction with the nuclei of atmospheric atoms (extensive air showers or EAS). The highest energy charged particle cosmic rays (E > 1017 eV) along with the ultra high energy gamma rays provide the most useful probes of the astrophysical sources from which they may originate. Particles of this energy will not significantly be affected by the intervening magnetic fields that they encounter, therefore, their trajectories should point directly back to their locations of origin[15]. The flux of UHECR cosmic rays is extremely low (10 -10 m-2·s-1). Collecting enough events to provide sufficient statistics for the determination of source locations will require increased exposure via increasing the area of ground based detector arrays, and/or increasing the field of view (of the sky) of detectors. The dearth of data also suffers from inherent bias due to the nonuniform exposure of experiments to the sky. Observations indicating the existence of large scale coincidences of EAS separated by distances of up to 100 km suggests interaction of UHECR with material in the interstellar medium[5],[14]. This data is sparse and does not at present represent statistically compelling evidence for large scale correlations due in part to the lack of high angular resolution in the determination of shower direction. The * [email protected]; phone 1 850 599-8385; fax 1 850 599-3557; www.neutralino.org PREPRINT v16mpf

Page 1 of 13

03/11/10 08:59:09 AM

expansion of such large area detector arrays such as AUGER to include both northern and southern hemisphere installations has the potential of full sky observation and increased statistics necessary for higher precision estimation of cosmic ray source locations[17]. In this paper, we describe the design (and plans for the deployment) of a scintillation detector sensor network which can form the basis of repeatable arrays of inexpensive detector nodes for increasing exposure to the UHECR flux and for higher-resolution determination of extensive airshower directions. 1.2 Existing Cosmic Ray Detection Technologies Ground based observation of UHE cosmic rays use the detection of air shower particles at ground level via scintillation and Cerenkov radiation in tanks or pools of water. Fluorescence due to cosmic ray excitation of atmospheric Nitrogen and Cerenkov emission by the secondary particle shower in the atmosphere may also be observed by telescopes designed for that purpose. Each approach has its benefits and limitations. Ground based scintillation and Cerenkov detectors can observe without regard to diurnal cycle or weather. On the other hand, air fluorescence and Cerenkov telescopes require good weather, clear skies and moonless nights for observation. The energy of the shower-initiating cosmic ray primary is usually determined by comparing signal data with Monte Carlo models of shower development in the case of ground based scintillation and Cerenkov detection. The observation of shower development in the atmosphere via fluorescence or air Cerenkov detection relies on calorimetric measurement of the energy deposited in the atmosphere in order to determine the energy of the shower initiating primary. The best approach is to combine both techniques where one acts as a check on the other as each suffers from different systematic uncertainties. The AUGER experiment does exactly this, combining both ground level sampling of the airshower and direct observation of the shower development in the atmosphere[13]. The AUGER experiment provides the largest exposure to the UHECR flux to date, about 30 events per year, and is limited to observation primarily of the southern sky. The AUGER North extension is planned to resolve this limitation. We hope to contribute to increasing exposure to the UHECR flux by means of mass deployment of relatively inexpensive networks of scintillation detector arrays. 1.3 Scientific & Educational Motivations of the Present Project A number of programs are underway to increase exposure of cosmic ray airshower observations involving institutions at all educational levels[19]. This paper describes the design of a scintillation detector sensor network for use at datacollection sites within the larger MARIACHI (Mixed Apparatus for Radio Investigation of Atmospheric Cosmic Rays of High Ionization) experiment, the details of which are described elsewhere [7],[8]. These efforts serve a broad set of goals that are both scientific and social in nature. The development of inexpensive cosmic ray detection systems will not only contribute to increasing the exposure necessary to increase observation statistics, but also may lower the financial barriers of participation in cosmic ray astronomy, in a way similar to the effect that the advent of widely available telescopes has had in optical astronomy, by fostering the creation of a large and enthusiastic network of amateur cosmic ray astronomers who may also contribute to the progress of science in cosmic ray astronomy.

2. SYSTEM DESIGN In this section, we give an overview of the overall architecture of our sensor network, and provide some additional details about the design of several key subsystems. 2.1 Sensor Network Architecture Figure 1 (below) schematically illustrates the overall architecture of our sensor network. A master repository operated by the MARIACHI project (c) collects all data that is gathered at individual data-collection sites, which may be scattered over an arbitrarily-wide area, anywhere from the scale of a metropolitan area up to globally. A single typical atmospheric cosmic-ray shower will impact an area several hundreds of meters across at ground level. So typically, we would want different data-collection sites to be separated from each other by several kilometers (at least) so that they will mostly capture independent shower events (events initiated by different atmospheric primary particles), thus effectively increasing the overall sensitivity of the wider network. However, sites that are more closely-spaced than this can still be useful, for improving the accuracy of data collected about individual showers. At each data-collection site (b) resides a small cluster of sensor nodes – nominally four, in our current design, although future versions may allow more – all positioned at about the same height and in the same orientation, and separated from

PREPRINT v16mpf

Page 2 of 13

03/11/10 08:59:09 AM

each other laterally by distances on the order of ~5-10m. The nodes might typically be located, e.g., at the corners of a medium-sized, squarish-shaped lab or classroom, mounted just below ceiling level, so that they will not be disturbed in the course of the room's ordinary use for other purposes. The relative 3-D positions of the detectors are measured by hand to ~cm precision (e.g. using tape measures), and the azimuthal orientation of the site (relative to geographic North) is characterized to within a few degrees, and its geographic location to within a few meters, using e.g. the property appraiser's diagram of the building's location and orientation on its plat. Each sensor node (a) consists of a scintillator block mounted in a shielded enclosure (to block ambient light & RF noise), which is monitored by a photomultiplier tube (PMT) connected to an FPGA-based front-end digitizer module (FEDM), which streams digitized PMT pulse data over a serial cable to a nearby wireless communication module (WCM), which then transmits the data over an ordinary 802.11b/g Wi-Fi network to a local “Central Server” PC (which is probably housed in the same or an adjacent room), which processes the data. This same PC also runs NTP (Network Time Protocol) in order to provide a local absolute time reference, and its clock is kept tightly slaved to a nearby NIST NTP server, providing us with roughly on the order of 10 ms precision in our absolute shower event time measurements. (In the future, we plan to improve our absolute time reference using a GPS clock to attain ~100 ns absolute precision.)

Figure 1. Overall architecture of the COSMICi sensor network. Each sensor node consists of a paddle-shaped block of scintillator material monitored by a photomultiplier tube with a custom base and Front-End Digitizer Module (FEDM); PMT pulse data is transmitted on arrival or at periodic intervals via Wi-Fi to a PC. Relative timing of local events to within 1 ns resolution is attained by reference to optical synchronization pulses that are beamed to each node in a room from the Central Timing Unit. Many geographically distributed sites (with various architectures) contribute data (in a common format) to the MARIACHI repository, for purposes of offline analysis.

Roughly at the center of the room (b), just below the ceiling, is an additional special apparatus called the Central Timing Unit (CTU), which periodically (every 409.600 microseconds, in the current design) emits an ultra-short, very intense 1 optical pulse from a high-flux red LED; this optical signal has a very short (~1-5 ns) rise time, and it is split and beamed in parallel to the four detector nodes along direct line-of-sight paths that have well-characterized (to within ~100 ps)

1 Around its spectral center of 635 nm, the peak spectral radiance of our pulsed LED was measured at roughly 15 kW/m2/sr/nm, comparable to the solar spectral radiance at these wavelengths of about 25 kW/m 2/sr/nm – although of course, the sun emits EM radiation over a much broader band, whereas nearly all of the LED's output falls within ±30 nm of its spectral peak. In terms of total radiance integrated over the entire EM spectrum, our LED's peak surface brightness is about 1/45 that of the sun. PREPRINT v16mpf

Page 3 of 13

03/11/10 08:59:09 AM

relative propagation delays. These pulses are then received (a) by sensitive optical detectors at each node, and are used to dynamically calibrate the relative local shower event time measurements to within <1 ns precision. The following subsections describe various subsystems of the above architecture in more detail. 2.2 Optical Synchronization Subsystem The optical synchronization subsystem, which consists of the Central Timing Unit (CTU) together with the optical sync pulse receivers at each node, is responsible for providing sufficient timing information to all nodes at the site to allow the relative arrival times of a given atmospheric shower's shock front at the various nodes to be identified with nanosecond precision. To this end, the central timing unit transmits a few-nanosecond rise-time optical synchronization pulse every 409.6 μs, which is received at each node and is used to align each node's idea of the absolute time, as well as to provide a definite reference point relative to which the arrival times of incoming pulses from the PMT sensor are measured. The synchronization interval of ~0.4 ms was chosen to be small enough so that the 1 ppm clocks on the nodes' digitizer boards would not be expected to drift out of their calibrated phases by more than 1 ns in between sync pulses, and large enough to allow our optical pulser to recharge fully between pulses (otherwise, pulse amplitude is reduced). The CTU (Figure 2) contains a custom Digital Timing Pulse Generator (DTPG) which uses a high-precision, 10.0000 MHz oven-controlled crystal oscillator (OCXO) from Connor-Winfield (part number BSOF3S3E), featuring 2 ppm frequency calibration and 10 ppb frequency stability, to clock a simple digital frequency divider circuit, which multiplies the clock period by exactly 4,096 and outputs square-wave, 100ns-wide pulses. This signal is used to trigger an IXYS Colorado PCO-7110-40-4 laser diode driver, which discharges a capacitor biased at a high voltage (up to 195 V) through a fast power MOSFET, producing a current pulse of up to 40 A with a 1-2 ns rise time and ~4 ns pulse width. Normally, this signal would be used to drive a laser diode, but, for safety and to reduce cost and increase the beam width (to simplify free-space beam alignment), instead we substituted a Nichia NSPRR70ASS High-Flux InGaN red LED, which we found is able to tolerate very high pulse voltages & currents (including pulses up to 100 V, 30 A in our experiments) when pulsed at a low duty cycle (we applied a duty cycle of at most 4 ns / 20 μs = 0.02% in our experiments), even though the LED's max rating is listed as 2.9 V, 200 mA (pulsed) in its datasheet. This may be because, given our very short pulse duty cycles, the average power dissipation in the part always remains in or below the neighborhood of its maximum power rating of 200 mW, so the amount of excess heating experienced by the part is minimal. (However, despite their already having survived several weeks' worth of continuous testing in the lab, it remains to be seen whether these LEDs will continue to function reliably on timescales of months or years even under these extreme conditions.)

Figure 2. Rough design of Central Timing Unit (CTU). Left: Side view of central apparatus. Right: Top view of optical platform. Power supplies & mechanical supports are not shown. The beam angle adjustment mirrors (right) can be manually rotated laterally as needed to aim the projected red spot (which, despite its ~0.001% duty cycle, is even visible to the naked eye when room lights are out) towards the active area of the optical sync pulse receiver located at each node.

The forward light output from the LED is then roughly collimated using a simple spherical lens (6.5 cm diameter, 6” focal length), which we found empirically yields a sufficient beam intensity at up to ~5 m distances to permit easy signal

PREPRINT v16mpf

Page 4 of 13

03/11/10 08:59:09 AM

detection. This beam is then split into four parts using a custom pyramidal shaped mirror (see Figure 2). A trio of semisilvered 50/50 beamsplitters could have been used instead, but our approach has lower cost. 2.3 FPGA-Based Time-to-Digital Converter (TDC) At each node, digitization of shower-pulse data from the PMT, as well as of time-synchronization pulses from the CTU, is carried out using a custom board (a.k.a. Front-End Digitizer Module, FEDM) developed by one of us (Junnarkar) and associates at Brookhaven National Laboratory. The basic principle of operation of this board, which has already been applied successfully in other particle-detection projects [10],[11],[12], is to use the LVDS (Low-Voltage Differential Signaling) inputs to an Altera Stratix FPGA as comparators that detect when each analog input signal crosses various reference threshold levels; the precise relative times at which different thresholds are crossed (in both directions) for each signal are measured to on the order of ~40 ps resolution by using two slightly detuned fast ring oscillators, which are reset by the threshold-crossing events, by simply counting the number of ring oscillator cycles that elapse before the relative phases of the two oscillators cross each other. By referencing sync pulses, board clocks, and ring oscillator cycles, all together, the precise times of pulse data points relative to each other within the local network can be defined to an overall precision of <1 ns, limited by rise-time uncertainties in the synchronization system. 2.4 Wireless Communication Subsystem Once the digital data packet describing a given PMT pulse has been prepared by the FEDM board, it transmits this data over a standard RS-232 serial interface using a UART. This basic serial protocol (at 115,200 baud) is sufficiently fast for our purposes, given the expected data rates (at most a few events per second at the target energy levels, depending on the detection threshold). The cable goes to an external Wireless Communication Module (WCM), which is currently a Laird EZURiO WISMC01BI Development Board. The central component of this board is an inexpensive ($100) stampsized daughter-card which can eventually be mounted directly onto the FEDM board in a later version of the system. A custom script running on this board (in Ezurio's UWScript language) maintains a TCP connection to a “Central Server” application running on the site's PC, and bridges the raw data from the serial port over this connection. The script also interprets commands sent back from the server to facilitate a variety of remote control and monitoring functions. 2.5 Central Server Software The Central Server software is still in the early stages of development. For coding this tool, we chose Python for its cross-platform portability, rapid development cycle, and its extensive available suite of open-source libraries, including support for TCP/IP, multithreading, cross-platform GUI programming (with the Tcl/Tk-based TkInter toolkit), etc. So far, we have only implemented a few simple server functions to monitor node status using virtual terminal sessions. Later, time permitting, we will implement a complete graphical user interface that will dynamically display the status of each node along with several visualizations of the incoming data, including graphs of the raw pulse waveforms, a 3D animated reconstruction of the shower wavefront's passage through the site, and a scatter-plot of the accumulated data points superimposed on a sky map, portrayed as colored circles showing the point of origin, energy, and reconstructed identity of the ultra-high-energy primary cosmic ray particle that is concluded to have produced a given atmospheric shower event. Of course, the raw and reconstructed data will also be preserved in a local database, and accumulated data will be periodically uploaded to the master repository maintained by MARIACHI, where it will be combined with data from other sites and archived for purposes of later offline scientific analysis.

3. ANALYSIS PLAN In this section, we summarize our plans for how the raw data collected by our sensor network will be processed and analyzed within our system to extract information of scientific interest. 3.1 Shower Direction and Cosmic Ray Primary Our system samples the EAS shower front at ground level allowing us to measure the angle, energy distribution and thickness of the shower front. The direction of the shower will be determined from the shower front angle in online analysis of the scintillation pulse waveform data. The waveform data is a record of time-tagged TDC values for each point of the pulse at five voltage thresholds (10 points per pulse). The shower front angle will be determined from the best fit of the shower front angle and front geometry (plane or spherical) to the time differences in the peak of the pulse PREPRINT v16mpf

Page 5 of 13

03/11/10 08:59:09 AM

from each detector in the array. The energy and type of the cosmic ray primary particle initiating the shower can be determined by fitting Monte Carlo simulations of our array to the data. We will use AIRES or CORSIKA codes to model the energy distribution in the showers of electromagnetic (photon) and hadronic (proton or nuclei) particle initiated showers and derive the expected pulse data for a shower of the inclination determined from the directional analysis. 3.2 Pulse Shape Model & Data Fitting Algorithm For now we are modeling each PMT pulse very simply as a dual-Gaussian peak, that is, one wherein the leading-edge and trailing-edge curves are assumed to follow normal Gaussian curves having the same amplitude and mean as each other, but in general different widths. These model pulses thus have four parameters (amplitude, peak time, leading edge width, and trailing edge width), so in theory, four data points (i.e., level-crossing times for two threshold levels) would suffice to determine a unique best-fit model in this class. To find the best-fit model, we start with the raw time-series data (which provides threshold-crossing times for each of the crossed thresholds) and we use a simple Newton's-method-based fitting algorithm, which empirically takes about 200 iterations to converge; on a modern PC, the convergence time should be short enough to permit real-time reconstruction of model pulses as they arrive at the central server. (In the future, we could even do the fitting in the front end FPGA, if it turns out that this may be useful to filter out spurious events on the basis of inappropriate pulse shapes.) After curve fitting, another important function of the central PC is to carry out coincidence detection, that is, filtering out those pulse detections that are not approximately replicated by near-simultaneous pulse detections received at the other nodes. Such isolated pulses can be generated by e.g. radioactive decays in the immediate environment, or by lowerenergy cosmic rays that are not constituents of a particle shock front resulting from a single dense, high-energy shower. 3.3 Shower Direction Reconstruction Once a shower event has been detected (near-simultaneous pulses from all four nodes), the next step is to reconstruct the shower energy and direction. The shower energy can be extrapolated from the pulse amplitude, which should be approximately the same for all four pulses as long as the scintillator blocks are identically oriented, since the shock front will be nearly planar on the scale of the site installation. Thus, in a dense high-energy shower, the portion of the shock front that passes through each scintillator should contain approximately an equal number of particles, propagating in the same direction relative to the scintillator geometry, and thus should produce light pulses of roughly equal intensity. The shower direction is reconstructed from the relative arrival times of the pulses at the four detectors. Actually, in principle, for near-planar shock fronts (small spherical sections) that are propagating at very close to the speed of light, as is the case in very high-energy showers originating from high altitudes, just three detectors would suffice to determine event time, azimuth, and elevation; but adding a fourth detector gives the potential for obtaining some information about the speed of the pulse (if less than that of light) and/or the depth in the atmosphere of the primary interaction. It also provides redundancy and an additional means to filter out spurious events. Our tentative algorithm for reconstruction/filtering is as follows. We begin with an arbitrary candidate event, identified in terms of the shower's original interaction coordinate (distance North and East of site location, and altitude), time of interaction, and velocity of wavefront propagation (initially assumed to be c). From these 5 numbers, the expected time of arrival of the shock wavefront at each of our 4 detectors can be straightforwardly calculated (since the locations of the detectors are precisely known). This can be compared with the actual times to determine the RMS error. A Newton's method fit can then be applied to iteratively adjust the event parameters to converge on a best-fit to the data. To prevent overfitting, we can impose logical bounds on the parameters, such as not allowing the shock front speed to drop below 0.9c, or above c, and/or not allowing the altitude to drift outside the middle-atmospheric range 10-120 km, where most of the primary cosmic-ray interactions occur. 2 (The accuracy of the altitude information so obtained will be negligible, anyway, since the distance between nodes in our present architecture is at least 1,000 times smaller than the altitude, so the effect of event distance on the relative arrival times will be very small and will tend to get lost in the noise.) 2 Strictly speaking, with the present geometry (all 4 detectors in a plane), there is a possibility of accidentally detecting and misinterpreting upwards-moving showers triggered by high-energy cosmic neutrinos that have passed through the Earth; if we wish to filter out (or select for) these events, then the fourth detector should be placed substantially outof-plane with the others (on the floor, say) which avoids the up-down mirror symmetry of the existing configuration. PREPRINT v16mpf

Page 6 of 13

03/11/10 08:59:09 AM

After the fitting process has converged to a best-fit within the bounds, we can then examine the remaining RMS error (which should be on the order of 1 ns or less) to determine if the data represents a viable candidate event. If not, then it is an accidental temporal coincidence of two or more independent events (i.e., not all from the same shower) at the four detectors, and we can discard it. If all the errors are within reasonable bounds, then we can keep the event. After the shower event time and location (relative to the site) have been determined, we can then compute its relative azimuth/elevation coordinates, and then (using the latitude/longitude of the observer location, as well as the current time and date), we can easily calculate right ascension & declination coordinates on the celestial sphere for the putative cosmic ray source (because these are very high-energy particles, which, even if charged, are not deflected significantly by magnetic fields). This absolute bearing information can then be easily mapped to galactic coordinates, to make it easier to identify any sources that are located in or near the plane of the Milky Way, or at the galactic core.

4. PRELIMINARY EXPERIMENTAL RESULTS At the present stage of the project, our demonstration prototype of the complete system is not yet complete, so we are unable to present results on the actual cosmic-ray data of interest. However, due to the uniqueness of our approach, the design and unit testing of the operation of various system components has nevertheless already involved a fair amount of experimental study. In this section we present some of the results obtained from these preliminary design experiments. 4.1 I-V characteristics of high-flux LEDs when pulsed at high currents In this set of experiments, we wished to examine how the effective resistance R = V/I of our LEDs varied when pulsing the LEDs at very high voltages, to determine if there was a point of diminishing returns (i.e., increasing differential resistance, below the maximum rated voltage of our pulser) where increasing the voltage further did not commensurately increase the peak current flow through the device (and thus its peak light output). If this turned out to be the case, then we might consider operating the pulser below its maximum rated voltage in order to achieve greater device reliability and lifetime (extended mean time to failure) without sacrificing very much in terms of output signal strength. To examine this matter, we systematically varied the input voltage from the high-voltage supply of the IXYS pulser, from 10 V to 195 V (the pulser's max rating) in 5 V increments - measurements at 5 V were skipped since at this level the output current peak was close to the noise floor, and could not serve as a reliable trigger for purposes of sample capturing on the oscilloscope. The role of this input voltage to the pulser is to charge up an internal capacitor whose stored charge is quickly switched across the LED when pulsing; this voltage thus directly impacts the peak applied forward bias voltage across the LED, as shown in the left graph in Figure 3 below. In this figure, the statistics for each data point were measured and accumulated over at least 1,000 pulses using a LeCroy WavePro 7200 digital oscilloscope. As we can see, the peak applied forward bias across the LED increases smoothly and almost linearly across the input voltage range, at a little less than half the supplied voltage. In an idealized version of the pulser's switching circuit , the peak voltage across the LED would simply be equal to the supplied voltage, but in reality, since the power MOSFET used in the pulser does not switch on instantaneously, the pulse-charge storage capacitor inside the pulser has already been partially discharged from its nominal voltage level by the time the current attains its peak level; and also, there is some additional voltage drop through the power MOSFET itself, due to its non-zero effective resistance. Additionally, at each voltage step, the peak LED output current was also measured using a current-monitoring resistor built into the pulser; since the peak current occurs simultaneously with the peak voltage, their ratio was used to derive the effective on-resistance across the LED at the voltage peak in each case. (This information was not already available from the device datasheet, since that only plotted IV curves within the LED's ordinary range of continuous operating voltages, below 5 V.) The peak current and effective-resistance results are plotted in the right-hand graph in Figure 3. The error bars shown on the resistance values were not measured directly, but were conservatively approximated using the worst-case assumption that the measured current and voltage fluctuations were perfectly anticorrelated. (In reality, they are probably highly correlated, and so the actual resistance fluctuations are probably much smaller than this.) This resistance curve, although it is fairly constant within a range of 2.7-3.1 ohms, has a number of statistically significant features which are currently not very well understood. There is a broad resistance minimum around ~38 V, a broad local maximum at ~65 V, and another apparent local maximum at 14 V. However, a more detailed study would

PREPRINT v16mpf

Page 7 of 13

03/11/10 08:59:09 AM

Peak LED forward current (A)

Peak LED Forward Bias Voltage (V)

100

50

40

3.5

35

3.0

30

2.5

25 Current (A)

2.0

Eff. R (ohms)

20 1.5

15

1.0

10

0.5

5 0

0 0

50

100

150

200

Pulser Supply Voltage (V)

0.0 0

10

20

30

40

50

60

70

Ef f ectiv e LED f orward resistance, ohms

be needed to ensure that these features are not measurement artifacts, and to explain them definitively in terms of device and circuit characteristics. In any event, the variation in the LED's effective resistance at high voltages is not large enough to justify any particular operating level below the maximum; as expected, the differential resistance is never negative. So, in our subsequent experiments, we normally powered the pulser at its highest rated voltage, in order to maximize the peak current and peak light output, so as to facilitate pulse detection. However, if, in the future we encounter reliability problems with the LEDs when pulsing them at these highest voltages, the information in this graph should be useful to help us predict how far we can turn down the applied voltage while still obtaining a detectable signal.

80

Peak LED f orward bias v oltage (V)

Figure 3. Voltage and current response characteristics of a Nichia NSPRR70ASS high-flux LED when pulsed at 2 kHz using the IXYS PCO-7110-40-4 laser diode pulser. The left chart illustrates how the measured peak forward bias voltage applied across the LED varies as a function of the voltage level supplied to the pulser's high-voltage power input. Ideally, this would be the identity function, but the power MOSFET used in the pulser does not switch on instantaneously, so the pulsecharge storage capacitor inside the pulser has already been partially discharged by the time the current attains its peak level. The right chart shows how the peak current through the LED (measured using a 0.1 Ω current-viewing resistor on board the pulser) varies as a function of the applied voltage, and plots the effective forward resistance of the LED (at the instant of peak voltage & current) over the range of peak voltages. The effective resistance varies in the range 2.7-3.1 Ω; its minimum level appears to be attained when the applied pulse peaks at 38V, corresponding to an 80V supply to the pulser.

4.2 Spectral characteristics of pulsed vs. unpulsed LEDs, including effects of device heating In our next set of experiments, we wanted to see whether and how the spectral characteristics of the LED's light output varied (if at all) in this high-voltage pulsed regime. In these particular tests, LED spectra were measured at short distances (1 cm) using an Ocean Optics Red Tide USB-650-VIS-NIR Educational Spectrometer. In our case, we did not deem it essential to calibrate the spectrometer for the wavelength-dependent variations in the sensitivity of its photoncounting pixels, because the pulse wavelength range was narrow (only about 15-30 nm) and this particular spectro meter's response is nearly constant across this small range. The background spectrum of the ambient room lighting (which was normal laboratory fluorescent lighting) was subtracted out as a reference spectrum, but was anyway fairly small compared to even the time-averaged spectral irradiance of the LED emission peak as seen at these short distances. The left part of Figure 4 (below) shows the spectra we obtained at pulse frequencies of 2 kHz, 25 kHz, and 50kHz (columns), at supply voltages of 50, 100, 150 and 200V (rows). The vertical scales and integration windows are the same in all these spectra, whose variations exhibit some interesting features, to be discussed below. For comparison, the right part of Figure 4 shows the LED spectra when a normal constant forward bias voltage is applied across it, for voltages in the range 2.0V-3.0V (the device's maximum rated voltage is 2.9V, and empirically, the threshold for producing a visible amount of light output is in the neighborhood of 1.5-1.6V). The height and area of these curves are not directly comparable to those at left, because of the difference in source distances (~2.5 cm vs. 20 cm). The actual spectral irradiance at 2.5 cm (and thus the vertical height of these curves) would be about 64 times greater than at 20 cm, and so these curves would not fit on the above chart. I.e., at ordinary constant voltage, the LED is much brighter on PREPRINT v16mpf

Page 8 of 13

03/11/10 08:59:09 AM

average than when it is pulsed, but this is not surprising, considering that the input pulse width that we use (4 ns) is so small. So, e.g., at a 2 kHz pulse frequency, the duty cycle of the device is only 0.0008%.

Figure 4. (Left) Spectra of pulsed LED at various combinations of supply voltages to the pulser and pulse frequencies. These spectra were obtained with the USB-650-VIS-NIR spectrometer at a 2.5 cm distance from the LED, with a 3 ms integration window. The vertical scale shows raw received photon counts from the spectrometer in arbitrary units, and is not calibrated for absolute spectral irradiance. The wavelength range shown in these figures is 550-700 nm. The spectral peak is normally in the vicinity of 636 nm, but shifts up to several nm longer wavelengths at higher voltages and frequencies due to heating – this is clear due to the timescale of tens of seconds for these shifts to occur after each change in operating conditions. Because of this, the device was allowed to equilibrate to its new temperature for on the order of 1 minute before data collection. (Right) Spectra of LED under normal constant-voltage operating conditions. These spectra were measured using the same spectrometer at a 20 cm distance from the LED, still with a 3 ms integration window. In each case, the peak output wavelength is in the general vicinity of 637 nm, but shifts up to several nm longer wavelengths at higher voltages.

Discussion of spectral features. Measurements on both sets of spectra are charted in Figure 5. In both constant-voltage and pulsed regimes, we noticed similar increases in the wavelength of the spectral peak at higher voltages, which may be attributed to the effect of increased temperature on the device's PN junction, causing a reduction in the semiconductor bandgap. This is a rather well-known and widely-reported effect (e.g., see[4],[20]), and the thermal nature of its origin is corroborated in our experiments by observing that the peak wavelength takes several seconds to converge to its new value after operating conditions are changed, suggesting a process of convergence to a new dynamic equilibrium thermal flow as temperature distributions in the device and its immediate surroundings adjust.Interestingly, although the LED is much brighter (by about 64×) when biased at a constant 3V than it is on average when pulsed even at the highest voltages and frequencies that we tested (200V, 50 kHz), the peak wavelength shift is actually greater in the pulsed case, suggesting that the device temperature (and thus, the average power dissipation) is greater in the pulsed case. To confirm this, we integrated instantaneous i·v through/across the LED over the pulse in the 200V, 50kHz case, and found that the average energy dissipation in the LED per pulse is 10.34 µJ; at 50 kHz the LED power dissipation is thus 517 mW, as compared with only about 450 mW when the bias is a constant 3V. This is unsurprising given the high peak currents that we pulse at, and the quadratic (P = I2R) dependence of power dissipation on current. What, then, accounts for the ~64× lower average brightness in the pulsed case? This is primarily explained by the fact that, at high pulse voltages, a larger part of each electron's energy loss as it crosses the device is dissipated ohmically as heat (e.g., in wires and p/n terminal wells), rather than getting emitted as light, since the energy of each emitted photon is predetermined by the (relatively stable) bandgap. A wavelength in the neighborhood of ~650 nm thus corresponds to an PREPRINT v16mpf

Page 9 of 13

03/11/10 08:59:09 AM

electron voltage drop across the bandgap of ~1.9 V; so, at a 75 V peak voltage drop between device terminals, only about 2.5% of each electron's energy is emitted as light, whereas at 3 V, about 63% of each electron's energy is emitted as light. This mechanism accounts for at least a factor of 25× of reduced emission efficiency. The remaining factor of ~2.6× may be accounted for in part by the reduced probability of radiative electron-hole recombination processes (as opposed to any of several possible non-radiative recombination mechanisms) for electrons crossing a high-tem perature junction, due to the increased energy and thus decreased density of momentum states at higher wavenumbers[16]. To summarize the interesting results from our spectral measurements, we found that the spectral emission peak was significantly wider in the high-voltage (195V) pulsed regime (30 nm FWHM) than with an ordinary, small forward bias (1.6V) across the same LED (14 nm FWHM), but the center wavelength was redshifted only insignificantly (by less than 1 nm) when the LED was pulsed at our nominal (2.4 kHz) operating frequency, indicating that despite the high applied pulse voltage, the effects of junction heating at these relatively low pulse frequencies were minimal.

Figure 5. Quantitative measurements of LED spectral characteristics in pulsed (bottom row) and constant-voltage (top row) regimes. These are based on the spectra in Figure 4. Peak wavelength is defined here as the wavelength at which the raw spectrometer count (which, over this narrow wavelength range, should be approximately proportional to spectral irradiance) is maximal. Peak height is the raw count of received photons (in a 1-nm wide bin) at the peak wavelength. Peak width is the full-width half-maximum (FWHM) measurement of the width of the spectral peak in nm. Average intensity is total spectrometer counts of photons received over the fixed 3 ms integration window. Note that the implied scale for the relative peak spectral radiance and total radiance of the sources (columns 2 and 4) is ~64× larger for the top row due to its ~8× larger distance from the source.

4.3 Optical Performance and Detector Response After characterizing the LED's IV and spectral characteristics, we wished to examine the overall performance of our endto-end emitter/detector system, including our beam-generating optics. Pulses were received using the Hamamatsu C5658 avalanche photodiode-based optical receiver module, placed at a distance of 16' from the beaming apparatus. But first, we wanted to confirm how the detector response varied with distance from the source. We found that with no help from any optics, at 15' distances, the free-space pulse signal from the LED was too small (barely above the noise floor of ±5 mV) to be reliably detected with the C5658 APD module. However, we found that insertion of a simple lens (6” focal length, 6.5 cm diameter) into the beam improved the situation considerably: With this lens, the captured portion of the beam was collimated well enough to produce a concentrated spot approximately 2” in diameter at 15' distance; when placed in this spot, the response of the C5658 was fully saturated at 1V, with a short (3 ns) rise time. For splitting this beam into four components to be directed at individual sensor nodes, we used the eMachineShop.com service to fashion a pyramidal polished Al mirror, shown in Figure 6 below. Unfortunately, in tests, we found that the imperfect degree of flatness and specularity of the pyramid sides led to a substantial increase in the beam divergence, enough to prevent the signal from being reliably detected above the 5-10 mV noise floor level at our target ~16' PREPRINT v16mpf

Page 10 of 13

03/11/10 08:59:09 AM

distances. Fortunately, we found that if a small (1” diameter, 1” focal length) convex light-gathering lens is oriented and positioned properly with the APD's active area at the focal spot, this is sufficient to restore the strength of the received signal to an oversaturated 1V amplitude (Figure 7, below). However, the optics at the receiving end will complicate the beam alignment process during site installation, so in future versions of this system, we may manufacture a new pyramid with improved mirror quality, or abandon the pyramidal mirror approach entirely, and revert to an earlier, more conventional design, which involved three plate beamsplitters arranged in a tree topology on an optical platform; this would resolve the beam divergence issue, and eliminate the need to precisely orient each receiver in addition to the CTU's steering mirrors.

Figure 6. Pyramidal polished Al mirror (base width 3”, sides inclined at 45°) as rendered in the eMachineShop design software (left), and as manufactured (right). A surface flatness tolerance of 0.02” was specified in the job settings; this reduced the manufacturing cost, but introduced a substantial amount of beam divergence. Fortunately, we are able to compensate for this using a light-gathering lens at the receiver. An alternative approach for splitting the beam would be to use three 50/50 beamsplitters in a tree configuration; this would take up more optical platform area, but would avoid the flatness problem.

5. CONCLUSIONS In this section, we sum up the present paper, including the project's current status, and other possible future applications. Current Status of Project. In this paper, we have outlined our design for a new distributed detector system for highenergy cosmic rays, consisting of multiple sensor nodes per site connected via Wi-Fi, and augmented with an unconventional nanosecond-precision free-space optics-based timing system, which permits reconstruction of the direction of origin of the primary cosmic ray particle initiating a given atmospheric shower. Most of the key components of the system have already been prototyped and tested, and preliminary experimental studies of the key optoelectronic subsystems used in our novel synchronization mechanism indicate that its non-standard design will nevertheless meet our specifications. The project is proceeding as planned, and the probability of a successful outcome is high. Other Applications. Our discovery that the Nichia NSPRR70ASS red LED, at least, can be pulsed at very high currents (up to at least ~30A) in ns-scale bursts at kHz frequencies without rapidly degrading, and that in this mode it achieves a solar-scale peak spectral radiance which is more than adequate for free-space, nanosecond-resolution optical pulse transmission over multi-meter distances even in ambient room lighting conditions, is interesting in and of itself, and may have other applications, for example in helping to synchronize microprocessor clocks in large multiprocessor computing architectures. This line of work may also motivate the development of a new class of LEDs that are designed more specifically for robustness under these extreme pulsed conditions. The approach described here uses a full digital TDC based architecture for digitizing the pulse data. The voltage threshold levels are programmable and can be the basis of arbitrary “triggers” for determining whether pulse data should be kept or discarded. Also, timing data can serve as software trigger in online or offline analysis for identifying events of interest beyond those of cosmic origin. This could extend the potential uses of our system to the observation of radiation

PREPRINT v16mpf

Page 11 of 13

03/11/10 08:59:09 AM

of terrestrial origin, either naturally occurring or from contraband nuclear materials, making the systems possibly useful as homeland security, defense and law enforcement tools.

Figure 7. Oscilloscope traces of LED current pulse (bottom) and receiver response (top). The high-voltage power supply of the IXYS pulser driving the LED is set at 195 V, the maximum recommended level. The horizontal timescale is 20 ns/division, and the vertical scale of both traces is 500 mV/division. The bottom trace measures the voltage across the on-board resistor that measures the LED current, at a ratio of 500 mV/10 A. The average peak voltage of 1.31 V shown in this measurement thus indicates a peak current of 26 A in this test, with a rise time of about 2 ns. The light pulse output by the LED passes through the beaming tower optics (lens and pyramidal mirror), crosses a space of 16 feet (almost 5 m) and encounters the 1” receiving lens and the C5658 APD module, which is powered at its maximum rating of +13.5 V. Careful positioning and alignment of the optical elements was carried out at both ends to maximize the response. The shape of the response pulse denotes that the APD module's output is substantially oversaturated at its maximum level of 1 V, indicating that even longer distances are in fact attainable. The total latency of 44 ns here is due primarily to cable delay in the receiver's scope probe.

ACKNOWLEDGMENTS This work is supported by the U.S. National Science Foundation (NSF) under the CREST program, award number 0630370. We also gratefully acknowledge donations of components and equipment received from Altera Corporation's University Program, and student internship support from the Florida Space Grant Consortium.

REFERENCES [1] Abraham, J., et al., “Observation of the suppression of the Flux of Cosmic Rays above 4 x 10 19 eV,” Phys. Rev. Lett. 101, 061101 (2008). Amsler, C., et al. (Particle Data Group), “Review of Particle Physics,” Physics Letters B667, 1 (2008). Beatty, J. and Westerhoff, S., “The Highest-Energy Cosmic Rays,” Ann. Rev. Nucl. Part. Sci. 59, 319-345 (2009). Bludau, W. and Onton, A., “Temperature dependence of the band gap of silicon,” J. Appl. Phys. 45(4), 1846-1848 (1974). Carrel, O. and Martin, M., “Observation of time correlations in cosmic rays,” Phys. Lett. B325, 526-530 (1994). Chi, X., et al., “The Origin of Ultrahigh Energy Cosmic Rays,” Ann. NY Acad. Sci. 647, 399-404 (1991). Damazio, D.O., et al., “A Simple and Cost-Effective Passive Radar Technique for Ultra High Energy Cosmic Ray Detection,” Nucl. Phys, B (Proc. Suppl.) 134, 217-219 (2000). [8] Damazio, D.O. and Takai, H., “The Cosmic Ray Radio Detector Data Acquisition System,” IEEE Nucl. Sci. Symp. Conf. Rec. 2, 1205-1211 (2004). [9] Greisen, K., “End to the Cosmic-Ray Spectrum?,” Phys. Rev. Lett. 16(17), 748-750 (1966).

[2] [3] [4] [5] [6] [7]

PREPRINT v16mpf

Page 12 of 13

03/11/10 08:59:09 AM

[10] Junnarkar, S.S., et al., “An FPGA-Based 12-Channel TDC and Digital Signal Processing Module for the RatCAP Scanner,” IEEE Nucl. Sci. Symp. Conf. Rec. 2, 919-923 (2005).

[11] Junnarkar, S.S., et al., “Next Generation of Real Time Data Acquisition, Calibration and Control System for the RatCAP Scanner,” IEEE Trans. Nucl. Sci. 55(1), 220-224 (2008).

[12] Junnarkar, S.S., et al., “FPGA Based Self Calibrating Time-to-Digital Converter for Time-of-Flight Experiments,” IEEE Trans. Nucl. Sci. 56(4), 2374-2379 (2009).

[13] Matthaie, G., “The AUGER Experiment Status and Results,” Astroparticle, Particle and Space Physics, Detectors and Medical Physics Applications, Proc. 10th ICATPP Conf., 229-238 (2007). [14] Ochi, N., et al., “Search for large-scale coincidences in network observation of cosmic ray air showers,” J. Phys. G. Nucl. Partic. 29(6), 1169-1180 (2003). [15] Prouza, M. and Smida, R., “The Galactic magnetic field and propagation of ultra-high energy cosmic rays,” Astron. Astrophys. 410(1), 1-10 (2003). [16] Schubert, E.F., [Light-Emitting Diodes (2nd ed.)], Cambridge U. Press, p. 55 (2006). [17] Sommers, P., “Cosmic ray anisotropy analysis with a full-sky observatory,” Astropart. Phys. 14, 271 (2001). [18] Sommers, P. and Westerhoff, S., “Cosmic ray astronomy,” New J. Phys. 11, 055004 (2009). [19] Wilkes, R.J., et al., “WALTA School-Network Cosmic Ray Detectors,” IEEE T. Nucl. Sci. 51(4), 1385-1388 (2004). [20] Yang, J.W., et al., “Selective area deposited blue GaN-InGaN multiple-quantum well light emitting diodes over silicon substrates,” Appl. Phys. Lett. 76(1), 273-275 (2000). [21] Zatsepin, G.T., Kuzmin, V.A., “Upper Limit of the Spectrum of Cosmic Rays,” JETP Lett. 4, 78 (1966).

[NOTE: The pagination in this preprint differs from that in the published paper due to the inclusion of headers/footers.]

PREPRINT v16mpf

Page 13 of 13

03/11/10 08:59:09 AM