Qualification of the CMS Barrel Pixel Detector Modules

Transcription

1 Preprint typeset in JINST style - HYPER VERSION Qualification of the CMS Barrel Pixel Detector Modules Sarah Dambach 1,2, Christina Eggel 1,2, Urs Langenegger 1, Andrey Starodumov 2 and Peter Trüb 1,2 1 Institute for Particle Physics, ETH Zurich, 8093 Zurich, Switzerland 2 Paul Scherrer Institute, 5232 Villigen PSI, Switzerland dambach@phys.ethz.ch, ceggel@phys.ethz.ch, ursl@phys.ethz.ch, Andrey.Starodumov@cern.ch, truebpe@phys.ethz.ch ABSTRACT: To be written. KEYWORDS: Particle tracking detectors, Detector alignment and calibration methods.

2 Contents 1. Introduction Barrel Detector Modules Controls of the Readout Chip and the Token Bit Manager Optimization Criteria Test Setup 9 2. Startup Tests and Threshold Measurements Startup Tests Threshold Measurements DAC Optimization, Tests, and Calibration DAC Optimization Functionality Tests Performance Tests Calibration Algorithms Production Results Production Verification of DAC Setting Test procedure Test suite I (after assembly) Test suite II (before mounting) Module grading Pixel defects Chip performance Module sensor quality Module production quality Overall production quality Summary of mounted modules Introduction The CMS experiment [1] is a general-purpose detector for the Large Hadron Collider at CERN. The CMS tracking system consists of two subdetectors: the silicon pixel detector at the center of the experiment and the silicon strip detector. The pixel detector comprises three barrel layers at radial 1

3 distances of 4.4cm, 7.3cm, and 10.2cm from the beampipe and two forward disks at ±34.5cm and ±46.5cm longitudinal distance from the interaction point. The barrel part consists of 672 modules and 96 half-modules, the forward part of 672 plaquettes. The tracking system of the CMS experiment is designed to provide precise and efficient measurements of charged particle trajctories and the reconstruction of decay vertices of long-lived particles. At the design luminosity of cm 2 s 1 on average about 1000 charged particles will emerge from the interaction region every 25 ns. This environment requires a radiation-hard tracking detector with high granularity and fast response. The readout chips of the CMS tracking detectors are fabricated in standard 0.25 µm CMOS technology, which is inherently radiation hard due to thin gate oxide and special design rules. XXX Also sensor radiation damage? XXX To keep the occupancy at the innermost layers below 1%, a pixelated detector is required. With a pixel size of 100 µm 150 µm in the rφ z directions the expected occupancy is at 10 4 in the innermost layer. This low occupancy allows fast track finding at the high-level trigger. In the barrel pixel detector, the drift of the electrons to the collecting pixel implant is perpendicular to the 3.8-T magnetic field. This results in a Lorentz drift spreading the charge over several pixels. Charge interpolation of the analog readout signal provides a substantially improved hit resolution compared to the single-pixel resolution. This paper is organized as follows. First a short description of the pixel barrel detector modules is provided, focusing on the aspects relevant for testing. In section 1.4 the test setup used for module qualification is discussed. Section 3 provides the implementation details of the individual tests performed in the course of module qualification, section 4 shows the results. The full test suite performed for module qualification is described in section 5. The scheme for grading the tested modules and the overall results for all barrel detector modules are given in section Barrel Detector Modules The pixel barrel detector is built in a modular way and comprises 672 modules and 96 half-modules with a total of pixels. A standard module has a size of 66.6mm 26mm, and weighs 3.5g, and comprises 16 readout chips (ROC). Half-modules, necessary to combine the two detector half shells without gap, contain 8 ROC. The barrel modules are built out of the following components (see Fig. 1). The silicon sensor with a thickness of 285 µm is micro-bumpbonded to the ROC by means of indium bumps with a diameter of about 20 µm [2], connecting each sensor pixel with a pixel unit cell (PUC) on the ROC. The sensor is covered by a High Density Interconnect (HDI), distributing signals and voltages to the ROC, and serving as an interface to the front end electronics. The connection is established over two cables: (i) a power cable for supply and bias voltages and (ii) a Kapton multi-channel signal cable for the control signals and the analog readout. The Token Bit Manager (TBM) chip on the HDI organises the readout of all ROC. The base strips beneath the ROC provide the necessary mechanical rigidity and are used to mount the module onto the support structure. On the detector up to 12 modules will form a Control Network. The purpose of the ROC [13] is to measure the charge produced in the sensor, amplify it, compare it to an adjustable threshold, store it during the latency of the L1 trigger, and finally to send its amplitude in zero-suppressed mode over an optical readout chain to the off-detector analog-digital 2

4 a) b) Figure 1. (a) Exploded view of a pixel barrel module. The components, from top to bottom, are: signal cable, power cable, HDI, silicon sensor, 16 ROC and base strips. (b) Photograph of a pixel barrel module. converters. Its amplitude will be referred to as pulse height (PH). The ROC consists of three main building blocks: 4160 pixel unit cells, 26 double column peripheries, and one control and interface block. The 4160 pixels are arranged in 52 columns and 80 rows in z and rφ, respectively. The readout is organized using the column drain mechanism [?], one token arranges the readout of two neighboring columns (called double column). 1.2 Controls of the Readout Chip and the Token Bit Manager A schematic view of the readout chain in the ROC is shown in Fig. 2. The module characteristics and performance are controlled by 26 digital-analog converters (DAC) and 3 registers on the ROC and three DAC on the TBM. The DAC on a ROC can be separated into two main categories: (i) DAC set to the same value on all ROC. While some of them are not used, the majority has been set to optimal values that do not differ from ROC to ROC. (ii) DAC adjusted for every single ROC. Both types of DAC are now introduced in their order along the readout chain. The detailed discussion of DAC optimization will follow in section 3. The DAC Vleak_comp is used to compensate for a possible leakage current in the sensor. At the startup of the experiment the leakage current is expected to be small, and this DAC is set to 0. A ROC-internal charge injection signal can be used to simulate deposited charge in the sensor. Its amplitude is controlled through Vcal, an 8-bit DAC with a high and low range setting. In the former one unit of Vcal corresponds to 65 electrons, in the latter to about 455 electrons (see section 3.4 for the calibration of these values). If not mentioned otherwise, Vcal will be given in high range DAC units. The DAC CalDel delays the charge injection with respect to the ROC clock, one unit in CalDel corresponds to XXXns. The two 4-bit DAC VrgPr and VrgSh and the two 8-bit DAC VwllPr and VwllSh control the preamplifier and shaper system; they are intended to be set to identical values, respectively. Figure 3a) shows the timewalk as a function of these four DAC. Here timewalk refers to the time difference of signals of different amplitude crossing the threshold. XXX Determination of timewalk XXX Figure 3b) shows the PH as a function of VhldDel for different values of VwllSh = VwllPr. Small DAC settings correspond to small timewalk, but also to small signal amplitudes. The 8-bit DAC VhldDel controls the sampling point of the signal amplitude. The 8-bit DAC Vhld- 3

5 Figure 2. Schematic view of the readout chain inside the readout chip. Del should be set so that the signal is sampled at its maximum value independent of its size, i.e., the timewalk should be minimal. Since the timewalk should be minimized and the PH maximized, an optimum was determined as VrgPr=VrgSh=0 and VwllPr=VwllSh=35. Figure 3c) shows the PH as a function of VhldDel for different low-range Vcal settings. The amplitude decrease for high Vcal settings is due to saturation in the readout chain XXX. A value VhldDel=160 is consistent with the 4

6 requirements above. a) b) c) Figure 3. (a) Timewalk as a function of the preamplifier and shaper DAC. For each measurement, VrgPr=VrgSh and VwllPr=VwllSh. (b) Pulse height (PH) as a function of VhldDel and VwllSh = VwllPr. (c) PH as a function of VhldDel for different Vcal values. The threshold is set for the entire ROC with the DAC (VthrComp). Four trim bits allow a threshold modification per pixel (scaled with Vtrim). These DAC depend on the desired threshold and have to determined for each ROC and pixel. After the PUC the signal is sent to the double column periphery where an offset can be added with VoffsetOp and the 4-bit DAC Vbias_sf. The two DAC have overlapping functionalities, and since VoffsetOp is optimized (see section 3.1), Vbias_sf =10 is fixed. XXX In a final step the signal level can be shifted (VIbiasOp, controlled by VIon and VOffsetRO, controlled by VIbiasOp) and in the control and interface block scaled (VIbias_PH). XXX There is an almost binary influence of the 8-bit DAC VIbiasOp on the PH: below VIbiasOp 20 no signal is seen independent of the amount of injected charge, while above this value the PH shape does not change. Therefore the setting can be done more or less arbitrarily, VIbiasOp = 50 was chosen. The 8-bit DAC VIon has a stretching influence on the PH. Since the pulse height range will be adjusted by VIbias_PH, it is set in the intermediate region: VIon = 130. The readout of all pixels is organized in two sequences: (i) a fast step (the signal passing through the comparator) to store the time of a hit and (2) a slow step (the signal passing through the sample and hold mechanism) to read the signal amplitude and the pixel address (column and row). For the first path every pixel on a double column sends a current to the periphery, its intensity 5

7 is adjustable by VIColOr. If more than one pixel was hit in a double column at the same time the currents are added. In the periphery a timestamp is created and stored in a time-stamp buffer. The second step is to read out the addresses and the signals stored in the sample and hold capacitances and to assign them to the corresponding time stamp. The 8-bit DAC VIColOr adjusts the amount of current which will be sent to the periphery. If the ROC is operated in the self triggering mode, a threshold on the number of hit pixels per double double column can be set. It is therefore important to know how large the current per pixel is. Since the self-triggering mode will not be used, the only concern is to have a current large enough that a hit pixel will be recognized, which is fulfilled for settings of VIColOr > 20; the default setting is at VIColOr = 99. The pixel address is sent from the PUC to the periphery as digital current levels and converted there into digital voltage levels. The threshold of this conversion is adjusted by the 8-bit DAC VIbias_bus, set to VIbias_bus=30. In the control and interface block the address levels can be shifted (Ibias_DAC) before they are prepended to the analog signal. Together they can be scaled (VIbias_roc) and will be sent out from the control and interface block. The 8-bit DAC VIbias_roc stretches the address levels and, likevibias_ph, the PH. Since VIbias_PH is optimized, VIbias_roc is set close to its maximum to keep maximal flexibility in adjustments. Different voltages have to be distributed over the ROC. Vdig and Vana determine the digital and analog voltages used in various ROC locations, VComp regulates the supply voltage of the comparator, and Vsf of the sample and hold circuit. Vdig is a 4-bit DAC and is used to regulate the digital voltage on the ROC. Since it does not affect the PH its only constraint is its influence on the address levels. It is set to 6, a value where the amplifier shows a linear behavior and the ROC-voltage is below the external voltage (2.5 V). The 4-bit DAC VComp regulates the comparator supply voltage. In the intermediate region of its possible range the comparator already works very reliably. Therefore it is set to 10. Irradiation may require that it has to be adjusted. VNpix and VSumCol are designed for adjusting the minimum number of hit pixels in a double column and the minimum number of double columns in the self-triggering mode of the pixel detector. Since this possibility will not be used, both of these 8-bit DACs are set to 0. The TBM [4] is controlled via three DAC. Dacgain stretches the digital TBM levels, Inputbias and Outputbias stretch both, the signals of the ROCs and and the TBM. Due to the high track density in the two inner layers of the barrel the TBM consists of of two parts with identical functionality. Normally only one half is used for controlling a module (single mode). In the two inner layers of the barrel detector the TBM is operated in the dual mode, each half controlling 8 ROC. Both Inputbias and Outputbias are 8-bit TBM DACs which have no influence on the PH if they are above a certain threshold (around 110). Both are set above this threshold to 128. All DACs are listed and sorted by category in Table Optimization Criteria Before describing the algorithms developed to optimize the performance of the ROC the possible optimization criteria and figures of merit are discussed. On the one hand, a reliable operation with minimal power consumption, yet maximum signal amplitude and minimal time walk, are criteria 6

8 Table 1. DAC and registers of the pixel modules. PUC refers to the pixel unit cell, CIB is the control and interface block, and the TBM is the token bit manager. DAC names in italic font indicate 4-bit DAC, DAC names in roman font indicate 8-bit DAC. Voltage regulators Vana, Vdig, Vcomp, Vsf PUC: Analog Vleak_comp, VwllPr, VrgPr, VwllSh, VrgSh, Vtrim, VthrComp, VhldDel Trigger VIColOr, Vnpix, VSumCol Calibrate VCal, CalDel Periphery VIbias_bus, Vbias_sf, VoffsetOp, VIbiasOp, VOffsetRO, VIon CIB Ibias_DAC, VIbias_PH, VIbias_roc Registers CtrlReg, WBC, RangeTemp TBM Inputbias, Outputbias, Dacgain quantifiable in the laboratory. On the other hand, parameters influencing the position resolution (determined from an ensemble of tracks) need to be validated through a detailed simulation and can be optimized only indirectly. The module readout is zero-supressed, i.e., only the PH and address of hit pixels are sent in analog form to the front end electronics. An illustrative example of a readout is shown in Fig. 4. The first eight clock cycles form the TBM header, followed by the readout of all ROC and terminated by the TBM trailer. The TBM header starts with three Ultra Black (UB) levels. An UB is a low level marking the lower bound of the analog signal range. The three UB are followed by a black (B) level defining the zero level of the differential analog signal. The four remaining clock cycles encode an 8-bit event number. The minimal readout of each ROC starts with an UB, a B, and a level called last DAC. This level displays the value to which the last addressed DAC was set, or the value of the temperature sensor. Each hit adds a block of six clock cycles to the analog readout, encoding the double-column (2 cycles) and row (3 cycles) of the hit pixel, and the pulse height (1 cycle). The addresses are encoded as digital signals on six levels. The readout is terminated by the TBM trailer, containing two UB, two B, and four clock cycles with the TBM status information. The analog readout is digitized in 12-bit ADC off the detector. To exploit maximally the available ADC range, the PH range should cover the same maximal range as the address levels. As the B level is the zero line, the range should be symmetric with respect to the B level. As the position resolution is mostly affected by the linearity of the PH in the low range, we dicuss now the quantification of the signal linearity. Figure 4b) illustrates a typical PH as a function of Vcal. For Vcal between 0 and 10 the signal is below the threshold of the comparator and no signal is visible. For 10 < Vcal < 50 the PH is very non-linear. This implies that the conversion of a specific PH (here around -700) to deposited charge is no possible (see section?? for a discussion of the influence on the position resolution). For 50 < Vcal < 125 the pulse height behaves linearly, above 125 it starts to saturate. This saturation does not pose a problem for position resolution since this corresponds to more than twice the minimum ionization charge. To quantify the amount of non-linearity in the lower Vcal range a hyperbolic tangent function is fitted to the PH: PH = p 3 + p 2 tanh(p 0 V cal p 1 ) (1.1) 7

9 a) b) Figure 4. a) Output of a module with one pixel activated on ROC 0. The initial eight cycles encode the TBM header, the following 6 show a single pixel hit on ROC 0, followed by 15 ROC without hit, and terminated by 8 cycles of the TB< trailer. b) Analog pulse height (PH) vs. Vcal The relevant parameter of this fit is p 1, which shifts the curve in the Vcal direction and can be used as a quantification of the non-linearity of the PH curve in the lower range. The smaller p 1 is, the more linear the PH curve is in the low range. a) b) c) Figure 5. Illustrations of aspects of signal linearity. Low-range linearity quantified with a tanh fit and pixel response considered a) linear and b) non-linear. c) Linearity in the full range. Figure 5 illustrates that values of p 1 1 imply a PH curve with a very linear behavior down to low Vcal values, while p means saturation in the low Vcal range. From visual inspection a value of p was chosen as target for the linearity, optimized through the DAC Vsf and balanced against the digital power consumption, as described in more detail in section 3. In addition to the linear behavior in the lower Vcal range, the linearity over a large part of the Vcal range and the full use of the entire available ADC range is important. To quantify this aspect, a polynom of 5 th degree is fitted to the PH curve and the tangent through the inflection point in the main region is determined. Starting from the inflection point in both directions, the PH-difference between the tangent and the fit is calculated. By definition, the linear range is determined as the region where this difference is smaller than 10% of the total pulse height range. The figure of merit is defined by a quadratic addition of the linear part in Vcal direction and the linear part in PH direction as shown in Figure 5c: linear range = ( V cal ) 2 +( PH) 2. (1.2) 8

10 1.4 Test Setup The testing of all barrel pixel detector modules was performed at Paul Scherrer-Institute (PSI) with the testing hardware setup described in this section. A desktop PC with a Scientific Linux 4 (SLC4) installation is the central control unit of the test setup. It is connected through USB cables to four electronics test-boards, specifically designed at PSI for testing the barrel pixel modules. Each test-board provides the necessary supply voltages and electrical signals (clock, trigger, readout,... ) for one module. To analyse the readout of the module, each test-board includes two 12-bit ADC sampling the analog signal in the interval [-2048, +2047], with 1 ADC unit corresponding to mv. The central control unit of the test-board is formed by a field-programmable gate array (FPGA) with an embedded processor. XXX Each module is connected through a module adapter board to its test-board. A Keithley high-voltage supply XXX provides the bias voltage (nominally 150 V, but ranging up to 600 V for leakage current tests) for the test-boards; one single output is fanned out to all four test-boards. The four modules are housed in a custom-built temperature-controlled box (TCB). The TCB allows a rapid temperature cycling at a well-defined humidity. All code and testing algorithms are implemented in a standalone C++ software package running under SLC4. For data storage and analysis the ROOT framework [5] is used. Fig. 6 shows an unified markup language diagram of the most important classes. The attributes and operations shown are only typical examples of the complete lists of variables and methods. Likewise the two test classes Trim and IVCurve serve as examples for the large number of implemented test classes. Furthermore, classes like those for the graphical user interface, the command line, the logging functionality, etc. are omitted for clarity. The core of the test classes consists of the six classes ControlNetwork, Module, ROC, TBM, DoubleColumn, and Pixel, representing the corresponding hardware entities. They reproduce their functionality like setting a DAC or enabling a pixel. These commands are sent to the test-board represented by the class TBInterface. The DoubleColumn and Pixel classes do not directly communicate with the test-board. All their commands require the specification of the ROC-ID, therefore they are processed via the ROC class. The TBInterface class itself makes use of the class USBInterface to send its commands to the (physical) test-board (at an earlier stage another interface was used for the communication between PC and test-board). The Test class provides common code to the derived test classes like Trim or IVCurve. An example for this common code is the function ModuleAction which loops over all ROC on a module and executes the test algorithm for this ROC. If this behaviour is not suitable for a derived class, as for instance in IVCurve, the derived class replaces it with its own code. The IVCurve uses the class Keithley to communicate with the high-voltage supply. The presence of the FPGA processor allows the execution of parts of the test algorithms directly on the testboard. This speeds up the tests by reducing the data transfer between PC and test-board. Especially interactive algorithms, where the test flow depends on the results of previous measurements, as is the case for threshold measurements benefit substantially. Table 2 shows a comparison of the duration of some test algorithms with and without running parts of the code directly in the FPGA. Without the FPGA processor the full test time would be longer by a factor of about three. 9

11 ControlNetwork modules GetModule 1 Module hubid DigitalCurrent Test testparameters ModuleAction 1 TBM tbmparameters SetRegister 16 ROC dacparameters rocid SetDAC Trim vcal AdjustVtrim IVCurve voltagestep ModuleAction TBInterface tbparameters SendCal 26 DoubleColumn doublecolumn EnableDoubleColumn Keithley port SetVoltage USBInterface buffer Write Read 160 Pixel trimbit EnablePixel Figure 6. UML diagram of the most important classes of the C++ testing software. Table 2. Test duration of different algorithms with and without running parts of the code directly in the FPGA. Remove this table XXX? Test duration per ROC [s] PC based code FPGA based code Trim Bits Test Bump Bonding Test Noise Measurements Trimming Pulse Height Calibration Startup Tests and Threshold Measurements This section discusses the low-level tests needed to ensure the basic functionality of the ROC and TBM. The threshold measurements are ingredients in most of the tests described in section Startup Tests The current test Provided by CE (?) xxx 10

12 The data trigger level (DTL) test determines the threshold below which an UB level is measured on the test-board. If the test-board detects three signal below this level, they will be interpreted as the UB levels of the TBM header and the ADC starts the sampling of the analog readout. It stops after detecting two signals below the DTL. The DTL is adjusted in the following way: First the DTL starting at zero is decreased until a valid analog readout is measured. Once this is achieved, the UB level is determined and the final DTL is set 100 ADC units above this value. The DAC-programming test is a simple check that all DAC of each ROC can be programmed: For each ROC the Vcal DAC is set to its extreme values 0 and 255 and the change of the last DAC is determined. If it lies below 20 ADC values, the ROC is considered to be defective. The TBM test checks the basic functionalities of both TBM channels. For both the communication is probed by reading out the event number. A second test concerns the readout mode of the TBM. The TBM can be read out in a single mode, in which the readout of all ROCs is sent to one analog channel. This is the default test mode and will be used for the modules in the third layer of the detector. In the dual mode, one half of the ROC is read out via the first analog channel, the other half via the second channel. This will be the readout mode of the modules in the first two barrel layers. The TBM test ensures that the module can be successfully operated in both modes by checking the length (i.e. the number of ROCs) of an empty readout. Before starting the other tests, the sampling point of the analog signal is optimized by adding a delay to the module clock with respect to the ADC clock. The sampling point is set to the center of the range, in which the pulse height is not more than 20 ADC units apart from the maximal value. The Pixel Readout To generate and read out a hit in a pixel, the following sequence of actions has to be taken: (1) Enable the double column of the corresponding pixel; (2) Enable the calibrate injection to the pixel; (3) Enable the readout of the pixel; (4) Send a calibrate signal to the module; (5) Send a trigger signal to the module. In the following the term to read out a pixel always refers to this procedure. 2.2 Threshold Measurements The threshold of a pixel can be measured in different ways. One possibility is to keep the threshold defined by the VthrComp DAC fixed and to find the Vcal value, at which a pixels start to respond. This kind of threshold will be called Vcal-threshold. The second possibility is to inject a signal with a fixed amplitude defined by the Vcal DAC and to find the VhtrComp value, at which this signal is above threshold. This threshold is referred to as VthrComp-threshold. Usually this measurement is done by reading out the hits in a fixed bunch crossing. This type of threshold is called in-time threshold. If a pixel has e.g. an in-time Vcal threshold of 60, this does not necessarily mean, that the pixel does not respond for Vcal values lower than 60. It only means that in the given bunch crossing no hits with lower Vcal values are registered. It is well possible, that by reading out the previous bunch crossing, signals with lower amplitudes can be observed. If a timing-independent threshold is required, the thresholds in different bunch crossings have to be measured. The minimum of all these thresholds is called the absolute threshold. If not mentioned otherwise, a threshold determination usually refers to an in-time Vcal-threshold. 11

13 The concrete measurement of a threshold proceeds in the following way. In a first step the considered DAC is varied in steps of 4 DAC units starting from one end of the DAC range. For each value the pixel is read out once. As soon as the response of the pixel changes, the scan stops and the current DAC value is returned. In a second step the threshold curve starting from this rough estimate of the threshold is measured with several readouts per pixels in steps of 1 DAC unit. If the readout efficiency reaches 50% the corresponding DAC value is returned. Due to the step size of 1 DAC value and the limited number of readouts, the precision of this measurement is 1.3 DAC values for 5 readouts per point. Since this threshold measurement has to be executed almost one million times per module during all tests, a more precise measurement by fitting the threshold curve is only done for the noise measurement (see below). 3. DAC Optimization, Tests, and Calibration 3.1 DAC Optimization Several DACs on the ROC have a big influence on its behavior, for example on the functionality or on the pulse height linearity. Their best setting varies quite much from ROC to ROC. Therefor they are dynamically adjusted for every single ROC. Dacgain The TBM 8-bit DAC Dacgain only has an influence on the analog levels of the TBM. Therefore it is the ideal candidate to set the ultrablack level of the TBM to a user-defined value, 1000 here. It is adjusted in such a way that the ultrablacks of both channels of the TBM differ least from the goal value, but lie below it. Since the position of the different levels is summetric around the black level and can not be shifted but only be stretched, this also fixes the position of all other levels, in particular the one of the highest TBM level to in this case. Ibias_DAC Ibias_DAC is an 8-bit DAC and has similar to Dacgain almost only an influence on the ROC levels. The little shifting influence on the pulse height can be ignored since this will be adjusted anyway in a later step. It is used to set the ultrablack of all ROCs to the same value as the TBM ultrablack. In the same way as for the TBM this also fixes the position of the ROC address levels. Ibias_DAC can be set upto a precision of 0.73 DAC units. Vana The 8-bit DAC Vana is set in such a way that the analog current drawn per ROC is 24 ma. The dependency of the analog current on Vana is shown in Figure 7. It can be set upto a precision of 0.55 DAC units. VthrComp versus CalDel All ROCs only work in a specific region of the VthrComp - CalDel range. To measure this region Vcal is set to 200 in low range DAC units (was found to be a good setting for many tests in [7]), five calibrate signals are sent for each pair of VthrComp and CalDel, and the number of readouts 12

14 Figure 7. Dependency of the analog current on Vana is counted. Since the working range does not change very strongly from pixel to pixel on the same ROC, this procedure is only done for one single pixel. A typical shape of the valid readout area is shown in Figure 8. Figure 8. Procedure of finding a stable working point in the VthrComp - CalDel space The used pair of the two DACs should lie as far away as possible from the the edges of the readout area. To find such a point in a first step the minimal value of VthrComp where a signal appears (horizontal line in Figure 8) is searched. From this point one goes up 50 VthrComp units and searches for the CalDel value in the middle of the readout range. This pair of the two DACs is defined as working point for calibration purpose. For the trimming of a ROC, VthrComp will be set in a different way. CalDel can be set upto a precision of 0.53 DAC units. VIbias_PH An important criterion of the DAC optimization is that the pulse heights and digital levels of all ROCs lay inside the same ADC range. In case of the levels this goal is already reached by setting the TBM and ROC ultrablack levels to a specific value. The only adjustment remaining is the one of the pulse heights, which should fill the goal ADC range. The general idea behind this procedure 13

15 is first to stretch or squeeze the pulse height range with one DAC and shift it afterwards to the desired region. As shown in Figure 9 with the 8-bit DAC VIbias_PH the complete pulse height curve can be stretched. Since this DAC has no influence on any address levels at all it is the optimal candidate to stretch or squeeze the size of the pulse height range to the favored one, 2000 (from 1000 to +1000) in this case. VIbias_PH can be set upto a precision of 2.62 DAC units. Figure 9. Influence of VIbias_PH on the pulse height curve Two DACs which only shift the pulse height curve and also have no influence on any address levels at all are VoffsetOp and VOffsetR0. Since they are correlated, they first will be discussed before coming back to the adjustment of the pulse height range. VoffsetOp versus VOffsetR0 VoffsetOp and VOffsetR0 are both 8-bit DACs which shift the pulse height curve and therefore have an influence on the linear range. The correlation between them is shown in Figure 10. It can be seen that for VOffsetR0 > 100 any linear range can be achieved by setting VoffsetOp correctly. Because of temperature and pixel to pixel variations of this behavior VOffsetR0 is set to 120 and VoffsetOp is adjusted afterwards. It can be set upto a precision of 0.41 DAC units. While the absolute value of the pulse height range is already adjusted with VIbias_PH, VoffsetOp can now be used to shift the pulse height curve in the goal ADC range. Since the variation of VoffsetOp also influences the pulse height range a little bit and VIbias_PH has a small influence on the position inside the ADC range, the procedure of adjusting those DACs needs to be repeated on average three times. Vsf The 8-bit DAC Vsf is the crucial DAC to get a linear behavior of the pulse height in the low Vcal range. The higher it is the more linear the pulse height curve gets, what is shown in Figure 11. The only problem is that the digital current of the ROC rises with increasing Vsf. Its absolute value depends on the difference between Vana and Vsf, whereas the former is already adjusted and will not be changed at this point anymore. The total digital current of a module as a function of Vsf of one ROC is shown in Figure

16 Figure 10. High range linearity in dependency of VoffsetOp and VOffsetR0 Figure 11. Influence of Vsf on the linearity of a pixel Figure 12. Influence of Vsf on the digital current of a module, Vana fixed The value of Vsf where the current starts to rise significantly is very chip dependent because Vana also varies from ROC to ROC. Beside the dependency on Vsf the linearity of the pulse height curve strongly depends on the temperature. As shown in Section?? the (non-) linearity of a pixel can be quantified by fitting its pulse height curve with a hyperbolic tangent function. The parameter p 1 of this fit is an indication for the linearity. To adjust Vsf it is increased in steps of five until this parameter is smaller than 1.4; if the increase of the digital current between Vsf = 0 and the recent setting is below 5 µa this setting will be used, otherwise Vsf will be lowered until the current increase is smaller. This optimization is done for an average pixel in terms of linearity. Vsf can be set upto a precision of 0.9 DAC units, while for time comsuming reasons it is set in steps of. VthrComp versus Vtrim The optimization of the 8-bit DACs Vtrim and VthrComp is part of the trimming which is described in Section

17 3.2 Functionality Tests Pixel Readout Test In the first part of this test, the functionality of the mask bit is checked. By enabling the mask bit of a pixel, the comparator in the PUC is disabled (cf. Fig. 2), therebye suppressing all hits in this pixel. This functionality is very important, since a noisy pixel can prevent a whole double-column from working properly by filling up the buffers in the double-column periphery. The mask mechanism is checked by enabling the mask bit and trying to read out the pixel. If a pixel hit is generated, the mask bit is defective. In the second part of the test, it is verified that sending a calibrate pulse to the enabled pixel, results in the corresponding hit information in the analog signal. For this, the pixel is read out 10 times with Vcal set to a value of 200 in the low range. If the hit does not show up in the analog signal all ten times, the pixel is called dead. Before this test the VthrComp and CalDel DACs have to properly adjust as described in 3.1. Trim Bits Test To individually adjust the thresholds of the pixels, each PUC stores four trim bits. By setting them appropriately the pixel threshold is lowered by an amount depending on the Vtrim DAC value. In the default untrimmed state, all trim bits are enabled, the corresponding trim value is 15. Disabling single trim bits will lower the pixel threshold. To test whether all four bits work as expected, the threshold is first measured in the untrimmed state. Afterwards all trim bits are enabled one after another and each time it is checked, that the threshold has decreases with respect to the untrimmed threshold. If the threshold difference is less than 2 DAC values, the trim bit is considered as defective. Figure 13 shows the threshold difference distributions for a ROC with no defects. For all pixels the threshold decreased by more than 10 DAC units. The Vtrim values used for the different trim bits are listed in Table 3. # Pixels 3 10 Trim value 14 Trim value 13 Trim value 11 Trim value Threshold Figure 13. Distributions of the threshold difference between untrimmed and trimmed pixels. For each of the curve one trim bit was disabled. 16

18 Table 3. Vtrim values used in the trim bit test. Trim value Vtrim DAC Address Decoding An individual pixel address consists of five clock cycles in the analog signal: two cycles encode the column index and three the row index [6]. Each clock cycle can take six different levels (c.f. Fig 4). To correctly decode the pixel address, these levels have to be well separated. To check this, the levels of all pixels in a ROC are measured and overlaid in a histogram as shown in Fig. 14. In this histogram, a simple algorithm searches for separated peaks. If exactly six of them have been found, the decoding limits are placed in the centres between two neighbouring peaks. These limits are used in the second part of the test, which records the analog readout of each pixel and checks whether the pixel generates the address which corresponds to its physical position on the ROC. # Pixels Analogue output level [ADC] Figure 14. Address-levels of all pixels in a ROC. The dashed lines are the separation limits used for the decoding of the pixel addresses. Bump Bonding Test An indium bump bond process for silicon pixel detectors has been developed at PSI [2]. To test the bump bonding quality, a fast algorithm using the possibility to send a calibrate signal through the sensor was devised [7]. The calibrate signal can either be injected directly to the preamplifier (using switch 1 in Fig. 15) or to a pad on the ROC surface (using switch 2 in Fig. 15). Choosing the 17

19 second option, the calibrate signal induces a charge in the sensor, which mimics a hit in the sensor pixel. Ideally, this hit is detected if the bump bond is present and not if the bump bond is missing. This ideal situation is deteriorated in two ways. Occasionally the bump bond is not completely missing but only has a poor connection to the sensor or the ROC. Even worse, for large enough amplitudes a hit is triggered although the bump is missing. These hits are supposed to originate from x-talk via a parasitic coupling between the calibration voltage line and the preamplifier. Based on this experience, the following algorithm was developed, to check the bump bonding quality. First the Vcal-threshold for the signal injection through the sensor is determined. Second the threshold for the parasitic x-talk is measured (i.e. with both switches open). The difference of the two thresholds (both measured in the high Vcal range) allows for a good discrimination between bonded and unbonded pixels. If the bump bond is missing, both thresholds are more or less equal, otherwise the difference amounts to 10 to 20 DAC units. It is found that the discrimination between good and bad bump bonds is better the higher the threshold (i.e. the lower the VthrComp value). This is made use of by setting the VthrComp DAC to a value, which is as low as possible, but still large enough in order to detect the pixel response due to the x-talk. The procedure has been validated by applying it to several specially prepared ROCs with sensors, from which a few bumps were removed manually before bump bonding. Fig. 16 shows the distribution of the threshold difference from which all missing bump bonds can be successfully identified. From the experience of many module tests, a bump bond was defined to be bad, if the threshold difference is larger than 5 DAC units. Figure 15. Sketch of the PUC components relevant for the bump bonding test. 3.3 Performance Tests Noise Measurements To identify noisy pixels, which potentially have to be masked, the noise of each single pixel is measured. The noise is determined by measuring the so called S-curve, which is the response 18

20 row Threshold [DAC units] column (a) # Pixels 3 10 good bumps bad bumps Threshold [DAC units] (b) Figure 16. Result of the bump bonding test for a ROC with known bump bonding defects. Figure (a) shows a map of the threshold difference to check the correct identification of the bad bump bonds. Figure (b) shows the threshold difference distribution, pixels with defective bump bonds are shown in red, good bump bonds are plotted in green. All pixels with missing bumps could be identified. efficiency of the pixel as a function of the amplitude of the calibrate signal. For an ideal pixel without any noise, this would be a simple step-function: zero efficiency below the signal threshold and full efficiency above. The effect of the noise is to smear out this step function. If the noise is assumed to be Gaussian, the S-curve has the shape of an error function, with a width proportional to the noise. A fast threshold scan provides a rough value for the threshold. In a window around this value the S-curve is measured with high precision, i.e. 50 readouts per point. The measurement is complicated by the fact, that the voltage of the calibration signal is not a monotonous function of Vcal. There are a few cases, where a higher Vcal DAC value corresponds to a lower voltage. The calibration voltage was measured as a function of Vcal for one ROC. This measurement is used to plot the efficiency directly as a function of the calibration voltage. These data points are then fit with an error function and the width and the position of the 50% point are extracted, see Fig. 17. The width is first converted to Vcal DAC units (1 Vcal DAC = 1.20 mv) and afterwards to electrons (1 Vcal DAC = 65 e ). With this procedure, the noise of a pixel can be 19

21 determined with a precision of 13 e. Efficiency Calibration voltage [V] Figure 17. S-curve fit with an error function to determin the noise of a pixel. Alternatively the noise can be determined from the pulse height measurement. For a fixed signal amplitude, the pulse height is measured 1000 times and the RMS of the resulting distribution is computed. Since other sources like the ADC, also contribute to the width of the distribution, the RMS of the black level distribution is measured and quadratically subtracted from the pulse height RMS. To compare the resulting number with the result from the S-curve method, it has to be converted into a charge. This is done with the help of the pulse height gain. Up to a mean offset of 20 e a good agreement between the two methods is found, therebye confirming that the S-curve width originates from the noise. Fig. 18 shows the result of both measurements for the same ROC. The edge pixel have a higher noise level due to their bigger size. Sensor IV-curve Defects in the silicon sensor are most easily found by measuring its leaking current as a function of the applied bias voltage. Problems with scratches or spikes would show up as breakdown at voltages below 100 V [8]. Since the sensor will be operated at voltages up to 600 V, the highvoltage is varied from 0 V to 600 V in steps of 5 V. The measurement of the leakage current is performed 5 s after the voltage has been set. The error on the current measurement estimated from repeated measurements of the same module is µa. The algorithm stops, if the voltage is either at 600 V or the leakage current exceeds 100 µa. The latter happens for many modules, but this is not considered as a problem if the breakdown voltage is above 200 V, since the behaviour of the sensor is expected to improve with irradiation. A typical sensor IV-curve of a good module is shown in Fig

22 Row Noise SC Row Noise PH Column Column 100 (a) S-curve noise (b) Pulse height noise # Pixels Entries 4160 Mean RMS Noise - Noise PH SC (c) Noise difference Figure 18. Pixel noise measured with two different methods: From the width of the S-curve (a), from the pulse height scattering (b), and the difference of the two methods (c). Leakage current [A] Voltage [V] Figure 19. Sensor IV-curve of a typical module at -10 C. 3.4 Calibration Algorithms Trimming Due to variations in the electronic components, each pixel has a slightly different threshold. To 21

23 achieve a uniform threshold for all pixels, the trim bits of each pixel have to be programmed to a suitable value. To do this, the trimming algorithm described below was developed. The only input parameter to the trim algorithm is the threshold (in Vcal units) to which the response of all pixels should be unified to. The three degrees of freedom which have to be properly adjusted are the VthrComp DAC, the Vtrim DAC, and the trim bit value of each pixel.vthrcomp sets the global threshold for the ROC and Vtrim determines, how much the trim bits lower this threshold. The influence of VthrComp and Vtrim on the threshold can be inferred from Fig. 20. Vtrim [DAC units] VthrComp [DAC units] Vcal threshold [DAC units] Figure 20. Vcal threshold of a pixel with all trim bits disabled as a function of VthrComp and Vtrim. Within the white area no threshold could be determined. The first step is to find the value of the VthrComp DAC which corresponds to the chosen threshold in Vcal units. This is done by measuring for each pixel its VthrComp-threshold. As for all other threshold measurements of the trim algorithm, the timing independent absolute threshold is measured. Since the thresholds can only be lowered afterwards, the minimum value of this distribution determines the global VthrComp value (a low VthrComp value corresponds to a high threshold). This value is used during the rest of the algorithm. As can be seen from Fig. 20 there is a maximal VthrComp value above which the ROC is not functional any longer. It turns out, that this limit has almost the same value for all pixels. In order to operate not too close to this limit, it is ensured, that the chosen VthrComp value is at least 10 DAC values apart from this upper limit. The second step of the trim algorithm is to determine an appropriate Vtrim value. To find this, the Vcal thresholds of all pixels are measured. The pixel which has the highest threshold is used to determine the necessary Vtrim value. For this pixel the trim value is set to zero and Vtrim is increased, until the threshold of the pixel is at the same level as the target threshold. The third step of the trim algorithm consists in setting the trim bits for all pixels. This is done by a binary search for the trim value, which gives a threshold as close as possible to the target threshold. The search starts with a trim value of 7 and comprises four iterations. At the end, all thresholds are measured once again to validate the procedure. For a not too low target threshold, the trim values fill the whole range from 0 (maximally 22

24 trimmed) up to 15 (not trimmed). The presence of the upper VthrComp limit poses a problem, when trimming for very low threshold is aimed for. In this case, the VthrComp DAC can not be set as high as desired in step one. This is compensated by a higher Vtrim value in step two, but at the price of a trim value distribution, which does not make use of the full available range. The trim value distributions after the trimming to thresholds equal to Vcal 20 and 60 are shown in Fig. 21. For the former threshold the trim values fill only the range from 0 to 12, therebye rendering the threshold distribution more coarse. # Pixels 3 10 Vcal 20 Vcall Trim value Figure 21. Trim value distribution after trimming to thresholds corresponding to Vcal 20 and 60. In the former case the trim values do not cover the full range, due to the upper VthrComp limit. With the final system it is not be possible to run highly interactive algorithms as the trim algorithms described above, because the detector control and readout systems (the FEC and FED) are separated. A solution to this problem is to use the trim parameters measured during the module qualification in the laboratory. During this procedure each ROC is trimmed to a threshold corresponding to 60 Vcal DAC units. It turns out that it is possible to extrapolate these trim parameters to any other threshold by parametrising them as a function of the threshold [10]. To obtain these parametrizations the trim algorithm was run for different thresholds. It is found, that VthrComp and Vtrim depend linearly on the threshold. The trim bits do not change significantly with the threshold. From the average of 16 ROCs the following parametrization was deduced: VthrComp(thr) = VthrComp(60) 0.65 (thr 60) Vtrim(thr) = Vtrim(60) 0.45 (thr 60) trimbits(thr) = trimbits(60) To validate the procedure, this parametrization was applied to a ROC and the resulting thresholds were measured, see Fig. 22. The measured thresholds deviate only little from the target thresholds. The widths of the threshold distributions are only slightly larger than those obtained by the full trim algorithm. 23