SVT Test Beam Analysis at UCSB


Index


Bad channels: correlations between testbeam data and calibration data

We have cross correlated the information between defects for the layer 2 module from a test beam analysis by Selenia Dittongo and calibration data taken in the lab at UCSB in April and July and in August at CERN. Note that the bad-channel-list was generated from calibration runs performed in April. The test-for-shorts had not yet been developed, so shorts were not explicitely flagged in the list. The results of the comparison are summarized here:

Pinholes
Inconsistencies between Selenia's analysis and the list of bad channels from calibration In conclusion, calibration data and beam data are in good agreement. The bad-channel-list to be used in the test beam analysis should be updated. We do not understand the behavior of the electronics for CNPs. This is mostly an academic question because (1) the rad hard design is different and, (2) the plan in production is to pluck the channels with pinholes. We also do not understand why a small number of shorts magically fixed themselves between April and August.


Offsets

Channel by channel offsets, obtained at the testbeam with the method described here, are given in the following four files: In this files there is one line per channel. Each line contains three entries:

Chip Number Electronics channel number Offset in DAC counts

The chip numbering convention here is that of the calibration program. For the layer 2 module, chips 0 to 6 are on the n side and chips 7 to 13 are on the p side. For the layer 5 module, chips 0 to 3 are on the n side and chips 4 to 7 are on the p side. The offsets are given in threshold DAC counts. To find the effective threshold for a given channel in a given run, you should subtract from the offset in DAC counts the threshold which was applied to the chip in that run (remember that the threshold DAC operates with negative logic, i.e. zero DAC counts corresponds to the highest possible threshold). The values of the chip thresholds should be accessible in the data stream. They are also summarized in four tables.


Time Over Threshold Calibration

Histograms and fits

We have extracted histograms of avearage TOT vs CAL_DAC for each channel, different shaping times, and for all thresholds used at test beam. HBOOK files containing these histograms are available in a gzipped tarfile. The file names are ttq_lx_y_z.rz, where NOTE: Due to a mistake in downloading the calibration constants, the data for the layer 5 p side at 200 nsec shaping time for a nominal threshold of 1.0 fC should be discarded.

The HBOOK histograms IDs are 900000 + 128 * chip_number + channel_number and the numbering convention is described above. Sample histograms are shown below (The CAL_DAC value is plotted on the x-axis, the average TOT value is plotted on the y-axis.




The histograms have been obtained from charge injection runs at fixed values of thresholds and have been individually fit to an empirical functional form. Noise hits have been rejected using the time stamp information.

Even in the absence of noise one expects some scatter in the TOT value for a given charge on a given channel. This is because even though the timing of the charge injection pulse is correlated with the system clock, the output of the shaper is sampled with a frequency lower than that of the system clock. Because of this scatter, it is necessary to collect several TOT samples in order to calculate the mean average TOT response as a function of charge.

Note that the TOT vs CAL_DAC shapes for the layer 2 and the layer 5 modules are different. The layer 2 module uses version 1 of the rad soft chip, while the layer 5 module uses version 2. There have been some changes in the shaper design between the two versions of the chip. The drop-off in average TOT for the layer 5 module at very high values of injected charge for the 100 nsec shaping time had also been seen by Al Eisner on the bench.

Note also that at 200 nsec shaping time the layer 5 response saturates the TOT dynamic range at high values of charge (recall that at the testbeam we operated the modules at 200 nsec with "no skip" in order to study the improvement in position resolution which could in principle be obtained with better TOT time resolution).

The histograms of average TOT vs CAL_DAC have been fit to the following functional form: where BREAK = 32 or 14 for the layer 2 and layer 5 respectively, and a,b,c,d,e are the five fit parameters.

These functions have been inverted to generated lookup tables of CAL_DAC as a function of TOT for every channel and for every nominal thresold used at the testbeam. These lookup tables are available in a
gzipped tarfile, and are intended to be used in the test beam analysis. The format of these lookup tables is described in a README file which is included in the tar file. A second set of lookup tables, based on the measured offsets, is also available, see below.

Channel Dependence

For a given nominal threshold, we expect channel-to-channel variations in the TOT response to charge injection due to
  1. Differences in the charge injection capacitors
  2. Differences in the gains
  3. Differences in the offsets
  4. Differences in the responses of the shaping circuits
It is easier to look at these differences using the TOT to CAL_DAC lookup tables, rather than the parameters from the functional fits, since these parameters tend to be highly correlated. The following figure shows the charge response as a function of TOT for the layer 5 module at 200 nsec shaping time for different channels at a nominal threshold of 0.8 fC. The charge has been obtained from the CAL_DAC value using nominal values for the charge injecting capacitor (65 fC) and the CAL_DAC to voltage conversion (1 count = 8.217 mV). For each value of TOT between 1 and 7, we loop over all channels and, using the lookup tables, we accumulate the corresponding charge into a histogram. We then show all seven histograms together (in different colors):




The fact that the histograms are not well separated means that using a channel independent lookup table as a function of shaping time and nominal threshold would result in a significant degradation of the intrinsic analog resolution. The effect on the position resolution is not clear without further study.

One of the most important reasons for the spread in response is probably due to the large spread in offsets. Using the
measured offsets it is possible to look at the spred in TOT response for channels with constant actual threshold (i.e. constant threshold-offset). This is shown in the Figure below, where we select channels with actual threshold within 0.5 threshold DAC counts of 0.8 fC (1 count = 4.29 mV. At 200 nsec shaping time the gain is approximately 83 mV/fC, therefore 0.5 counts correspond to 0.025 fC).




The separation between histograms has improved considerably. This suggests that we should also try to generate lookup tables which are just functions of the actual threshold and the shaping time.

Preliminary lookup tables based on (threshold-offset) have been generated and are available in a gzipped tarfile. These lookup tables have been generated (for the different chip versions and different shaping times) by combining all of the calibration data, and by taking the two-sigma-truncated mean response of all channels as a function of threshold-offset. Since no smoothing or fitting was done, the lookup tables are still a little ragged near the near the edges where the statistics are poor.


Problem with Threshold ?

Unfortunately the discussion above is not entirely consistent. The plot for Actual Threshold 0.8 fC indicates that TOT=1 corresponds on average to something like 0.5 fC. This is significantly lower than the 0.8 fC threshold. This could be due to a shift in offset (remember, Actual Threshold = (threshold-offset)), or to something else. Note that the gain/offset and TOT calibration data were taken on different days.

In order to try to understand what is going on, we can look in detail at the calibration data for one channel on the Layer 5 module, p-side, shaping time = 200 nsec. The gain/offset calibration was based on two threshold scans at fixed charge injection with CAL DAC values of 2 and 3 (corresponding to approximately 1 and 1.5 fC). The error function fits are shown below.




Unfortunately this was done only once, so we cannot directly compare these curves with curves taken at a different time. However, by taking data from different TOT calibration runs we can obtain a (small) set of points that can be directly compared with those above. This is shown in the next figure.




These three plots show the hit fractions for the same channel from the TOT calibration runs compared to expectations based on the parametrizations obtained from the threshold scans used in the gain/offset calibration (these are the dashed curves). The comparison is made for CAL DAC = 1,2, and 3. We actually did not directly measure the parametrization for CAL DAC = 1. What is shown in the third plot is the expected parametrization for CAL DAC = 1 based on the CAL DAC = 2 and 3 parametrizations and assuming linearity in the response. Note that not only were the dashed parametrizations obtained on a different day than the day on which the points were taken, but also not all points were taken on the same day.

A shift in offsets would effectively result in a shift (to the left or to the right) of the dashed line predictions. The CAL DAC = 2 and 3 points agree fairly well with the parametrization, but since most of them are so close to 100% they would also be consistent with a parametrization that is significantly shifted to the left (for example). On the other hand, the CAL DAC = 1 points are in poor agreement with the parametrization. Better agreement could be obtained by shifting the parametrization to the left.

The higher than expected fraction of hits at CAL DAC = 1 is directly responsible for the lower than expected threshold inferred from the TOT response curves.

One may worry that the CAL DAC = 1 response could be seriously biased by noise hits. However, the time stamp information shows that the hits are correlated with charge injection.

It is possible that at very low values of charge injection (i.e. CAL DAC = 1) the system does not respond as expected. It turns out that this was never checked at the UCSB
Test Bench. We will check this soon.

We have also looked for higher statistics evidence for or against the hypothesis of an offset shift. This is done in the following way on a channel-by-channel basis:





There is a systematic difference between eff1 and eff2. The sign of this difference is consistent with an offset shift of the same sign as the shift needed to explain the TOT = 1 inconsistency.


Useful Test Beam Links


Claudio Campagnari Page Last Updated: Oct 29, 1997