RECENTLY, the Silicon Photomultiplier (SiPM) gained

Transcription

1 2009 IEEE Nuclear Science Symposium Conference Record N28-5 The Digital Silicon Photomultiplier Principle of Operation and Intrinsic Detector Performance Thomas Frach, Member, IEEE, Gordian Prescher, Carsten Degenhardt, Rik de Gruyter, Anja Schmitz, and Rob Ballizany Abstract We developed a fully digital implementation of the Silicon Photomultiplier. The sensor is based on a single photon avalanche photodiode (SPAD) integrated in a standard CMOS process. Photons are detected directly by sensing the voltage at the SPAD anode using a dedicated cell electronics block next to each diode. This block also contains active quenching and recharge circuits as well as a one bit memory for the selective inhibit of detector cells. A balanced trigger network is used to propagate the trigger signal from all cells to the integrated timeto-digital converter (TDC). Photons are detected and counted as digital signals, thus making the sensor less susceptible to temperature variations and electronic noise. The integration with CMOS logic has the added benefit of low power consumption and possible integration of data post-processing. In this paper, we discuss the sensor architecture and present first measurements of the technology demonstrator test chip. I. INTRODUCTION RECENTLY, the Silicon Photomultiplier (SiPM) gained interest as a potential candidate to replace photomultiplier tubes for reasons of ruggedness, compactness and insensitivity to magnetic fields. Other advantages of solid state detectors in general are their low operating voltage, low power consumption and large scale fabrication possibilities. Today, silicon photomultipliers operate in an analog way. The passively-quenched Geiger-mode cells of the SiPM are connected in parallel through a long interconnect, and the resulting output signal is therefore the analog sum of the individual currents of all cells. Hereby, the good intrinsic performance of the SPAD is not fully utilized, as the generated signal is deteriorated by the relatively large parasitics of the on-chip interconnect, the bond wires and the external load. Furthermore, susceptibility to electronic noise and high sensitivity to temperature variations are typical characteristics of the analog SiPM. Also from the system perspective, large scale applications of analog silicon photomultipliers imply some design challenges. In systems comprising several tens of thousands of channels, reading out every single channel can become a difficult task. Usually, a dedicated multichannel mixed-signal ASIC is needed to condition and digitize the silicon photomultiplier output signals. As the single photon response is still in the mv range, the signals can be easily affected by interference, Manuscript received November 13, The authors are with Philips Digital Photon Counting, Weisshausstrasse 2, Aachen, Germany (telephone: , digitalphotoncounting@philips.com, web: Fig. 1. Scintillation light detector systems based on the analog (a) and digital (b) silicon photomultiplier. electronic noise or unstable baseline due to high dark count levels, thus rendering single photon trigger impossible. Interference from the switching digital part into the low-noise analog front-end of the readout ASIC can additionally affect system performance. Differential current-mode logic design can be used to minimize the switching noise in the digital part of the ASIC at the expense of significant increase in power consumption and heat generation. A typical singlechannel sensor/readout system based on the analog silicon photomultiplier is shown in Fig. 1(a). The same functionality can be realized in a single chip according to the scheme shown in Fig. 1(b). Here, the SPADs are integrated with conventional CMOS circuits on the same substrate. Each SPAD has its own readout circuit, which also provides means for active quenching and recharging of the SPAD. A one bit memory cell integrated next to the SPAD can be used to selectively enable or disable the respective diode. Each cell, composed of the SPAD itself and the corresponding electronics block, is connected to the time-to-digital converter via a configurable, balanced trigger network. A separate synchronous bus is used to connect each cell to the photon counters to determine the number of detected photons. Eventually, correction look-up tables and other data post-processing could be implemented on the same chip. Also, the bias voltage generation could be fully integrated with the /09/$ IEEE 1959

2 Fig. 2. Microphotograph of the test chip. sensor and a digital feedback control loop could be used to automatically adapt the bias voltage to the actual value of the breakdown and excess voltages, making the device insensitive to temperature drifts and process variations. II. SENSOR ARCHITECTURE Like its analog counterparts, digital silicon photomultiplier pixels consist of arrays of Geiger-mode microcells, each capable of detecting single photons. Contrary to the conventional SiPM, however, each cell is capable of detecting and storing exactly one photon. Upon the detection of a photon, the avalanche is actively quenched using a dedicated transistor, and a different transistor is used to quickly recharge the diode back to its sensitive state. SPAD breakdown results in an immediate voltage change of approximately the excess voltage at the anode. The shape of the anode voltage pulse follows an exponential defined by the capacitance and the inner resistance of the diode, as long as the quenching transistor is left open. Upon reaching the inverter threshold, the anode voltage is forced to the breakdown voltage level by closing the quenching transistor, thereby stopping the current flow through the diode. The combination of the diode capacitance and the quenching transistor feedback ensures proper storage of the information. The quenching transistor is disconnected and the diode is reset to sensitive state by a separate recharge transistor. Each cell provides a fast asynchronous trigger signal and a slower synchronous data output signal. The trigger signal is connected to a balanced, low skew trigger network connected to the on-chip time-to-digital converter (TDC). The trigger network can be configured to start the TDC at the first photon or, alternatively, at higher photon levels to reduce the number of triggers in the case of high pixel dark count rate. Except for the first photon trigger level, any diode breakdown is assumed to be a dark count and is automatically reset if it does not lead to a trigger within 10 to 15 ns from its occurrence. This embedded refresh logic prevents dark counts from accumulating and, eventually, reaching the trigger threshold. The data output signals of all cells in a column are connected to the same data line and the data is applied to the data line using a row output enable signal. Then, the acquired data can be read out by selecting individual rows of the array one after the other and reading the data lines at the periphery of the pixel. To simplify the design of the chip, the pixel is composed of four identical subpixels, each consisting of an array of cells. The cell size is 52 μm 30 μm with 50% fill factor including the cell electronics. One diode has been omitted in each subpixel to make space for the trigger logic at the center of the pixel, thus resulting in 2047 sensitive cells per subpixel. Each subpixel has periphery logic on two adjacent sides to control the access to the cell inhibit memory and to provide the means to readout the data and recharge the diodes. An integrated time-to-digital converter is connected to the trigger block at the pixel center via the master trigger line. A microphotograph of the technology demonstrator test chip is shown in Fig. 2. The TDC is located at the right hand side of the pixel array and has a block size of 96 μm 3300 μm. We designed diodes with 20, 25 and 30 V nominal breakdown voltage. All measurements presented in this paper were obtained using a sensor with 25 V nominal breakdown voltage. The actual bias voltage of the sensor equals the breakdown voltage plus the 3.3 V excess voltage and is typically in the 28 to 30 V range, as the breakdown voltage of the sensor can differ from die to die due to process variations. A. Trigger and Validation Thresholds As mentioned in the previous section, the main driver behind subdividing the pixel into subpixels is the complexity reduction of the full-custom design. As a side-effect, the data readout time is also reduced by a factor of two, and the implementation of the various trigger thresholds is simplified. The trigger thresholds are implemented in fully digital way using logic operations on the subpixel trigger lines. Assuming that ST 1,... ST 4 are the first-photon trigger signals corresponding to the subpixels 1,... 4, the master trigger MT is implemented using one of these operations: MT =1 = ST 1 ST 2 ST 3 ST 4 MT 2 =(ST 1 ST 2 ) (ST 3 ST 4 ) MT 3 =(ST 1 ST 2 ) ST 3 ST 4 MT 4 = ST 1 ST 2 ST 3 ST 4 This way, the trigger network can be configured to start the TDC at the first photon or, alternatively, at 2, 3, 4 photons. The TDC is stopped by the next rising edge of the system clock, which is used as a global time reference. A coarse counter connected to the system clock is used to extend the time measurement interval to ns. The implication of this triggering scheme is that, except for the first photon trigger, the trigger thresholds are statistical thresholds, as, for MT 2, the second photon can fall on the same subpixel as the first photon and so forth. Higher trigger (1) 1960

3 trigger no yes READY VALID? COLLECT READOUT RESET 5-40ns ns 320ns 10ns Fig. 3. Typical acquisition sequence. global pixel recharge and TDC reset, and then back to the READY state. In the technology demonstrator test chip (Fig. 2), the pixel state machine has been implemented in an FPGA and the corrections have been done offline in software. This approach offered greater flexibility in testing and comparing different acquisition sequences and simplified the debugging process. In the final sensor the acquisition controller will be placed next to the pixel on the same substrate. levels allow to control the system dead time at the expense of a slight loss of time resolution. This is especially important in applications where reduction of the dark count rate by active sensor cooling is difficult to implement. A second, higher-level energy-like threshold is implemented in a similar way to distinguish valid events from dark counts. The pixel is subdivided into 64 mutually exclusive regions to allow the adjustment of the validation threshold to up to the minimum of 64 photons. The validation logic makes use of the fact that many photons are detected at the very beginning of the scintillator pulse, thus increasing the probability of the selected regions to detect a photon. On the other hand, the probability to detect a dark count event in every selected region in this short time span is close to zero, as dark count events are generated independently of each other. The validation signal is tested at a user-defined time after the trigger has been detected and a fast pixel reset is issued in case of a dark count event to minimize sensor dead time. B. Data Acquisition A typical data acquisition sequence is shown in Fig. 3. The pixel state machine starts in the state READY with all diodes charged above their breakdown voltage and recharge transistors open. The master trigger starts the acquisition sequence, forcing the pixel controller to change the state to VALIDATE. The pixel controller stays in this state for a userdefined time of 5 to 40 ns. After the validation hold-off timer expired, the validation signal is checked to determine if the event is indeed a real light pulse or a dark count. In case of a dark count generated event, the pixel state machine changes to RESET to quickly reset the pixel and change back to the READY state in preparation for the next event. In case of a real scintillator pulse, the validation threshold is reached and the state machine changes to the state COLLECT. While in this state, the pixel waits for the scintillator pulse to decay. The collection time is user-defined between 5 and 2560 ns. The photons impinging on the sensor are detected and stored in the cells for later readout. After the expiration of the timer, the pixel state machine goes to the state READOUT. In this state, each line of the sensor is selected separately and the number of photons detected in the line is added to the photon accumulator. While reading out one line, the preceding line is recharged to avoid large current surges during the global reset of the pixel. As the sensor is still sensitive during the readout, half of the readout time contributes to the collection time. Finally, the pixel controller goes to the RESET state for C. Saturation Correction The photon counting capability at low light levels is realized in the same way as in the analog SiPM, i.e. by spreading the photons in the light pulse over the cells of the sensor. As long as there are many more cells than photons, the deviation from a linear sensor is negligible. At higher photon counts, however, the saturation cannot be neglected anymore and (2) can be used to calculate the number of photons detected by a sensor having N cells active cells. This equation is valid only in the case that each cell cannot detect more than one photon during acquisition, which is the case in the digital silicon photomultiplier. ( ) N detected = N cells 1 e PDE N photons N cells (2) Equation (2) can be readily solved for N photons, and a lookup table (LUT) can be used to correct the number of detected photons for sensor saturation. As the number of active cells can be different from one pixel to another, one correction LUT per pixel must be provided. Area permitting, this LUT as well as the TDC calibration LUT could be integrated on the same chip, thereby simplifying the overall system architecture. III. DARK COUNT RATE The possibility to selectively activate individual cells enables detailed characterization of basic sensor parameters like the breakdown voltage, sensitivity, time resolution and trigger network skew, and dark count rate. For example, to measure the dark counts of a single cell, only the desired cell is enabled and the trigger level is set to first photon trigger. The dark count rate can be easily measured by counting the triggers per second. This procedure can be done for all diodes in the pixel successively. The result is the dark count rate (DCR) map as shown in Fig. 4. It is evident that only very few diodes contribute significantly to the total dark count rate. Switching these diodes off can significantly reduce the total dark count rate of the pixel and thereby minimize the system dead time. Fig. 5 shows the histogram representations of the dark count rates of all diodes in the pixel at four different temperatures. The distribution of the dark count rates for most of the diodes is close to a Gaussian with approximately 90 to 95% of the diodes having a typical dark count rate close to the average. Only the remaining 5 to 10% of the diodes show abnormally high dark count rates due to defects. The average dark count rate of a good diode at 20 C is approximately 150 cps. This can be reduced by an order of magnitude by cooling the sensor to 0 C and, in the extreme case, to less than 1 cps per diode 1961

4 Fig. 6. Total dark count rate of the sensor at different temperatures. Fig. 4. Typical dark count rate map of the sensor. dark count rate of the pixel for the four different temperatures is shown in Fig. 6 as a function of the resulting relative PDE. In this particular pixel, switching off 5% of the cells would reduce the dark count rate of the pixel by a factor of two. Higher factors have been observed in other pixels, sometimes leading to a reduction of the dark count rate by more than an order of magnitude. Fig. 5. DCR histograms of all SPADs in a pixel at different temperatures. below -45 C. Having the possibility to measured the statistical distribution of the dark count rates of all individual diodes is also a powerful tool for process development and monitoring, as any systematic defect distribution becomes easily detectable as a deviation from the Gaussian distribution. The total dark count rate of the pixel is the cumulative sum of the dark count rate distribution, shown in Fig. 5, starting at the entry with the lowest dark count rate and ending at the user-defined maximum. As switching off diodes is equivalent to loss of sensitive area, the pixel dark count rate can also be computed as a function of the number of active cells or, in other words, the resulting relative photon detection efficiency (PDE) of the pixel. Based on the data from Fig. 5, the total IV. PHOTON DETECTION EFFICIENCY Photon detection efficiency (PDE) describes the probability to detect a photon of a certain wavelength. The factor includes the quantum efficiency of silicon, the avalanche generation probability, and the fill factor of the sensor. While the quantum efficiency of silicon is very high in the visible region of the spectrum, the actual charge collection depends strongly on the internal design of the avalanche photodiode. The avalanche generation probability denotes the likelihood of an carrier attracted into the high-field region of the junction to start an avalanche eventually leading to the junction breakdown. This depends on the electric field profile and the carrier type, as the ionization coefficient for electrons is different from the one for holes. Finally, the fill factor heavily depends on the lateral design of the device and the area taken by the cell electronics. In some applications, the PDE loss due to the fill factor can be mitigated by using microlenses. However, this is not possible in typical scintillator crystal readout as the light emission is more or less isotropic. Therefore, high fill factor is key to high overall PDE of the sensor. Increasing the cell size also helps to increase the fill factor at the cost of earlier saturation of the pixel. Fig. 7 shows the PDE of the test chip, measured using a monochromator setup and a calibrated reference diode. First, the response of the optical system was measured using the reference diode, and the photon flux at the position of the sensor was calculated. Subsequently, the reference diode was replaced with the test chip and the flux was attenuated with a calibrated neutral density filter. The spectral response of several individual cells was recorded and corrected for the dark count background of the diodes. As the individual cells 1962

5 Fig. 7. Photon detection efficiency of the sensor at 3.3 V excess voltage. Fig. 8. Dark count map used in the randoms correction. TABLE I EFFECTIVE PHOTON DETECTION EFFICIENCIES FOR COMMON SCINTILLATORS. Scintillator Effective PDE (%) Light yield (photons/kev) LYSO CsI(Na) CsI(Tl) NaI(Tl) BGO LaBr 3 (Ce) of the pixel were measured sequentially, any distortion due to optical crosstalk was excluded from the measurement. All data was measured at room temperature, using the nominal excess voltage of 3.3 V. The measurement range was limited by the setup to nm. Table I lists the expected effective photon detection efficiencies for a number of scintillator materials. The effective PDE is obtained through the convolution of the device PDE with the normalized scintillator emission spectrum. The typical light yield of the scintillator according to [1] is also given in the table for reference. V. OPTICAL CROSSTALK Optical crosstalk is caused by photons generated during the avalanche discharge of the diode [2]. These photons can initiate a secondary breakdown in one of the neighboring diodes. The probability of photoemission is directly proportional to the current density in the high-field region of the junction. Thus, reducing the current density inside the diode helps to reduce the optical crosstalk. In the analog SiPM, this typically leads to the reduction of the gain, thereby reducing the signal to noise ratio. In the digital SiPM, the current is reduced by discharging the diode through a dedicated quenching transistor. Additionally, the crosstalk can be further suppressed by optical isolation between the diodes. The flexible sensor architecture enables direct measurements of the optical crosstalk. The basic idea is to find a diode with a high dark count rate in a relatively quiet environment. This Fig. 9. Optical crosstalk of center cell to its neighbors. diode can be used as a light generator to generate optical crosstalk to its neighbors. Then, the crosstalk can be measured selectively for all combinations of the light generator and all of its neighbors. To find a suitable generator diode, we first measure the complete dark count map of the sensor. Subsequently, the program sorts all diodes according their dark count rates and checks the dark count rates of the 5 5 array surrounding the potential light source. If the dark counts of all its neighbors are below a user-specified limit, the actual measurement is started. The measurement starts with the re-acquisition of the dark counts of the 5 5 field, as the conditions may have changed since the measurement of the full pixel DCR map. The dark counts of the individual diodes are needed later to correct the measured data for randomly coincident dark counts. The result of the measurement is shown in Fig. 8. In a second step, the light generator diode and one of the test diodes are activated simultaneously with the trigger level set to first photon and validating all events. The received data packets can contain only one or two photons per event. Photon count = 1 indicates a dark count in one of the two diodes. Photon count = 2 means either a random coincidence of two dark 1963

6 Fig. 10. Histograms of the photon counts for attenuated laser pulses. Fig. 11. Time resolution at low photon counts. counts, or optical crosstalk between the two diodes. The true rate of optical crosstalk can be determined by subtracting the expected randomly coincident dark counts using the previously measured DCR map (Fig. 8). This procedure is carried out for all diodes of the 5 5 test field separately. The resulting optical crosstalk between the light generator diode and all its neighbors is shown in Fig. 9. VI. LASER MEASUREMENTS A picosecond laser (36 ps FWHM, λ = 410 nm) was used to characterize the photon counting performance of the test chip. The defocussed laser beam was attenuated to flux levels of one to fifty photons per pulse, and one million pulses were acquired for each attenuator setting. The trigger threshold was set at the first photon level and all events were validated. The average number of photons in the pulse for each measurement is derived from the histograms shown in Fig. 10. The data was acquired at 3.3 V excess voltage and -20 C to suppress the dark count offset due to the finite collection time of the sensor. The measurement was repeated at room temperature with essentially the same results except for photon fluxes below ten photons per pulse, where accumulated dark counts became visible as a fixed offset of the number of detected photons. The time resolution for each attenuator setting was obtained from the same data set. The time resolution as a function of the mean number of photons in the laser pulse is shown in the Fig. 11. The time resolution follows the 1/N relationship predicted by theory [3]. The only exception is the first point, which shows an improved time resolution. The root cause of this anomaly is currently being investigated. The contributions of individual system components to the time resolution of the sensor were investigated separately at low and high photon fluxes. The contribution of the TDC to the time resolution is 20 ps full-width at half-maximum (FWHM). Under low light flux conditions, the SPAD contributes 54 ps FWHM mainly due to the avalanche spreading uncertainty. Negligible jitter but significant systematic skew of 110 ps FWHM have been found in the trigger network. Manual fine- tuning of the wire lengths is underway to eliminate the trigger network skew in future designs. VII. TEMPERATURE SENSITIVITY One drawback of the analog SiPM is its pronounced sensitivity to temperature variations. Temperature affects the ionization coefficients of the electrons and holes in silicon [4]. This leads to a temperature drift of the breakdown voltage of the diode. Assuming constant biasing conditions, any change in the breakdown voltage V bd leads to a proportional change of the SiPM gain G, according to G = C (V bias V bd ) (3) q with C being the diode capacitance including any parasitics, and q the electron charge. In the digital SiPM, the voltage level at one of the SPAD terminals is sensed and digitized by an logic gate. Therefore, the digital SiPM is insensitive to any change in the breakdown voltage as long as the switching threshold of the gate is reached. The remaining drift observed in the digital SiPM is due to the change in the photon detection efficiency, caused by the temperature-dependent avalanche initiation probability. This drift can only be compensated for by adapting the bias voltage of the device. One advantage of the digital SiPM is the possibility to integrate the bias voltage generator on the same die. A variable bias voltage source could be realized using e.g. a charge pump circuit. The charge pump, together with a suitable temperature sensor or direct breakdown voltage measurement circuit would allow to completely compensate any temperature dependencies of the photon detection efficiency of the device. It would also simplify system integration, as the breakdown voltages of individual devices can differ significantly from each other. We measured the drift of both PDE and TDC using an attenuated picosecond laser with approximately 2100 photons per pulse. The nominal time resolution at this photon count was 24 ps FWHM. We observed a drift of 0.33% per degree Celsius in the average number of photons per pulse over a 1964

7 Fig. 12. Temperature drift of the mean photon count in the laser pulse. IX. SUMMARY We have developed a fully digital implementation of the Silicon Photomultiplier. The device is manufactured in a high-volume CMOS process and includes the single photon avalanche photodiodes, the detection and readout circuits as well as the time-to-digital converter on the same chip. The sensor is capable of detecting single photons with a time resolution of 140 ps FWHM. At higher photon counts, the time resolution of the device increases to 24 ps FWHM. The device has low temperature dependence and low power consumption. Measurements using LYSO scintillator crystals demonstrate the intrinsically good performance of the sensor. Energy resolution of 10.7% and a coincidence resolving time of 153 ps FWHM were measured under normal operating conditions. One of the benefits of the presented sensor architecture is its flexibility and extensibility. Depending on the area and power constraints, advanced data processing can be integrated on the same die, thereby opening the way to detector-onchip designs. Integrated data reduction algorithms and the possibility to daisy-chain sensors can further simplify the construction of large-area detector systems. Moreover, the various configuration options enable the user to optimize the detector performance for the given operating conditions. ACKNOWLEDGMENT The authors thank Dr. Hein Valk of NXP Semiconductors for many constructive discussions during the process development and CMOS integration of the single photon avalanche photodiode. Fig. 13. Temperature drift of the pulse time stamp. range of 15 Cto25 C (see Fig. 12). This drift is an order of magnitude lower compared to the analog silicon photomultiplier [5]. As the time-to-digital converter is integrated next to the SiPM pixel, both the TDC and trigger network are also affected by any changes in the die temperature. The time stamp drift has been measured to be 15.3 ps per degree Celsius (Fig. 13). This drift can be measured and compensated for using a dedicated electrical SYNC signal, which is intended to synchronize and calibrate the individual time stamping circuits in large detectors. REFERENCES [1] Saint-Gobain Crystals, Physical Properties of Common Inorganic Scintillators. Available: [2] D. Renker and E. Lorenz, Advances in solid state photon detectors, Journal of Instrumentation, vol. 4, no. 4, p. P04004, [3] R. F. Post and L. I. Schiff, Statistical limitations on the resolving time of a scintillation counter, Phys. Rev., vol. 80, p. 1113, [4] S. M. Sze, Physics of Semiconductor Devices, 2nd ed. John Wiley & Sons, [5] C. Degenhardt et al., The digital silicon photomultiplier - A novel sensor for the detection of scintillation light, Nuclear Science Symposium Conference Record, J04-1, VIII. SCINTILLATOR MEASUREMENTS The performance of the sensor with scintillators of different sizes has been measured and published in [5]. As an example, two3mm 3mm 5 mm LYSO scintillator crystals were measured in coincidence using a 22 Na source. The crystals were wrapped in Teflon tape and mounted to the sensors using MeltMount. The measurements were done at room temperature and at the nominal excess voltage of 3.3 V. The energy resolution was found to be 10.7%, and a coincidence resolving time of 153 ps for events in the photopeak was measured. 1965