WorksPhysics and Astronomy
Testing and Characterization of Wafer-Scale MAPS Prototypes for the ALICE ITS3 Upgrade
Abstract
The ALICE experiment will upgrade the innermost three layers of its vertexing detector, the Inner Tracking System (ITS), during the next LHC Long Shutdown (LS3) with a novel, bent, ultra-light MAPS-based tracker. Six wafer-scale sensor chips will be bent into three cylinders, held in place only by carbon foam, leaving no material except for the silicon die in most of the ALICE central barrel acceptance. Two prototype ASICs, approximately in length, called MOSS (MOnolithic Stitched Sensor) and MOST (MOnolithic Stitched sensor with Timing), have been produced. These two chips follow complementary approaches to evaluate the use of stitched CMOS sensors for the first time in an HEP experiment. This article gives an overview of powering tests, functional studies, pixel matrix characterization, and in-beam tests of both test structures. The overall yield of MOSS is measured to be approximately per region (1/80th of a chip). This number takes into account powering, as well as functional aspects such as digital and analog pulsing. Two major failure modes have been identified and understood: short in the power grid of the chip and readout issues, that can be clearly attributed to the readout architecture design. Disregarding these issues, the overall yield increases to about per region. Furthermore, it is shown that MOSS can operate with efficiency and fake-hit rate up to TID and NIEL. The MOST prototype successfully demonstrated use of power gating, which allows for disconnected parts of the pixel matrix from the power grid in case of shorts. MOSS and MOST successfully proved that designing stitched wafer-scale sensors is feasible and deliver valuable input for the design of the final ITS3 ASIC.
Introduction
In the context of the ALICE ITS3 upgrade project [1], a second submission (Engineering Run 2, ER2) has been designed in a 65 nm process. The key feature of this submission are two large scale stitched sensor prototypes. The MOnolithic Stitched Sensor (MOSS) [2] and Monolithic Stitched Sensor with Timing (MOST) measure 259 mm in length instrumenting a large fraction of the 300 mm wafer in a single chip. A picture of the actual wafer is presented in Figure 1. MOSS and MOST follow complementary design approaches in various aspects, which are highlighted in the following.
Figure 1. Wafer of the ER1 submission. The smaller boxes show instances of the babyMOSS and babyMOST made up of a single RSU and both end caps.
The MOSS prototype measures 14 mm in width, while the MOST prototype is 2.5 mm wide. The pixel pitch of MOSS is 22.5 m in the upper half and 18 m in the lower half, while MOST employs a single pitch of 18.5 m. Both chips share the basic concept of layout, where the central part housing the pixel matrix is a Repeated Sensor Unit (RSU). This RSU is placed 10 times along the long axis of the chip, where interconnection is achieved by stitching. Both short ends of this strip are terminated with separate designs. One of these end caps facilitates the readout of the chip, although the MOSS can optionally also be read out through pads on the long edge of the sensor, which provides a backup in case of faults in the data transmission backbone along the chip. MOST has been designed with an integration density at the maximum dictated by this CMOS process, whereas the MOSS is intentionally designed less dense to possibly assess the impact of integration density in case of insufficient yield.
The powering approach is different in both prototypes. The MOST pixel matrix is powered by two separate analog power domains and a digital power domain, that are distributed globally across the full chip. As for large chips the probability of having at least one short is significantly higher potentially lowering the overall yield to unacceptable levels, the MOST implements granular power switches as a mitigation strategy. These switches allow to detach groups of 256 pixels (analog) or 352 pixels (digital) from the global power grid allowing for normal operation in the remaining chip. However, issues in the power grid or the switches itself can not be mitigated in this way. To address potential switch-related issues, the MOSS power grid is organized with increased granularity, with each half-RSU featuring three power domains electrically isolated from the rest of the chip.
The diode biasing of MOSS is realized in a standard way, where a negative voltage can be applied between the substrate and the front-end ground. This is illustrated in Figure 2a. An alternative approach was chosen for MOST, where the lower rail of the front-end is shifted up by a voltage with respect to the global ground, creating a potential difference across the diode. In order to maintain the operational margin of the front-end, the analog supply voltage AVDD also needs to be increased above the nominal 1.2 V. Shifting back to the regular voltage level is performed in the comparator. A schematic is given in Figure 2b.

Figure 2. Biasing schemes of both large-scale ER1 chips.
The readout of MOSS is strobed and the pixels possess a hit-latch, which is read out via a priority encoder and additional digital logic in the chip periphery. The MOST is not strobed and purely hit-driven. Upon activation of a pixel, its address is streamed onto one of four global column readout lines shared by all pixels in the column. While this approach enables direct Time-of-Arrival (ToA) and Time-over-Threshold (ToT) measurements, it also introduces a potential risk: if two pixels fire within a short time interval, their signals may collide on the shared readout line, resulting in a sequence that cannot be uniquely decoded.
Along with several other small chiplets, the areas outside of the central rectangle of the wafer are filled up with babyMOSS and babyMOST chips, which consist of single RSUs of their respective large versions (see Figure 1). The design of the RSU and the end caps is identical to the ones used in the large-scale chips.
Powering Yield
For the final ALICE ITS3 a low density of critical defects is essential. Conventional detector modules allow for picking the best working chips for assembly. In the ITS3 approach, selecting individual chips is not possible and it is necessary to find six wafers for which a sufficient fraction of units in the center of the wafer are operable. To roughly estimate the yield that can be expected for the final chip (produced in the same 65 nm technology) a mass testing campaign was carried out. A total of 120 MOSS chips underwent a procedure of cross-impedance measurements between all power nets, followed by slow power ramping under supervision of a thermal camera and another impedance measurement. It is found that thermal hotspots appear during power ramping, which can be correlated with an increased supply current in one of the power domains. In some instances, at a given voltage the current sharply drops to a lower level and continues to rise slowly with increasing voltage normally. In these cases a change of impedance before and after the ramping is observed. Furthermore, the majority of thermal hotspots correlate spatially with the power grid, which leads to the overall hypothesis that shorts are existing in the power grid and that there is a certain probability that a given short is removed (burn-through) by supplying a certain voltage. Importantly, these hotspots occur over all of the chip and not only at the stitching boundaries, indicating no problem with the stitching process. Figure 3a shows the distribution of tested half-RSUs. Every half-RSU is categorized into: OK-I – the half-unit can be powered normally, OK-II – half-unit can be powered normally, but a transient current was observed, and LIMIT – a given half-unit can not be powered with a current limit of 500 mA, where the normal operating currents are in the order of a few tens of mA. In total, only approx. 4.3 % of half-RSU show persistent shorts and are not operable with the given current limit. All other half-RSUs can reliably and repeatedly be powered, even after burn-through [3].

Figure 3a. MOSS yield based on 1620 half-RSUs. Taken from [4].
Figure 3b shows a similar distribution for MOST, where OK-I chips show a current below 50 mA and OK-II chips exceed 50 mA. LIMIT means that a given chip exceeds a current of 100 mA. A normally working chip draws approximately 10 mA per power domain. Due to the global power supply, a tested unit of the MOST is a full chip, which is about 2.3 times as large in area as a MOSS half-RSU. 56.1 % of MOST chips can not powered. Naturally, this number is higher due to the increased chip area. Due to the presence of switches between the power grid and the pixel matrix itself, failures occurring already during power ramping with open switches indicate problems in either the power grid or the switches themselves. The former is compatible with the findings of the MOSS testing campaign as the general design of the power grid is similar.
Figure 3. Powering yield of both large-scale ER1 test structures.
Figure 3b. MOST yield based on 66 full chips.
Functional Yield
The functional yield of MOSS has been extensively studied in a mass testing campaign of 82 chips from 14 wafers. Testing steps are as follows:
Powering: Apply power to a unit, apply clock signal, perform reset, set nominal operating point.
Digital periphery: Write to and read back from digital registers.
Analog periphery: Assess linearity of biasing DACs.
Pixel matrix readout: Assert hit latches, then data readout. Afterwards analog pulsing.
Pixel matrix performance: Threshold and noise uniformity evaluation and fake-hit rate measurement.
Figure 4 illustrates the achieved yield per region, where one region is a quarter of a half-RSU. From about 86 % that are passing the powering stage, only about 0.2 % (0.1 %) of all tested units fail digital (analog) periphery tests. A larger fraction of about 8.1 % of all tested units are failing the pixel matrix readout test. However, the dominant mechanism of this failure is pixels in which the hit latch can not be deasserted. This problem is specific to the simplistic readout architecture of the MOSS prototype and such failures are not transferrable to a final chip. The pink bars show the achieved yield where regions with these issues are not considered in calculation. Nevertheless, about 0.6 % of all regions are failing this step due to > 1 % dead or noisy pixels, which are in principle transferrable to future chips. Finally approximately 0.3 % of regions show insufficient threshold or noise uniformity or an excessive fake-hit rate in > 1 % of pixels, giving a total yield of about 76.4 %, or 84.7 % if aforementioned readout architecture issues are removed from the calculation. As discussed in § 2, the main mechanism of failure for units that can not be powered is understood and will be mitigated. When not considering the powering issue and all issues specific to the readout architecture a total yield of about 98 % (normalized by region) is achieved [4]. The overall functional yield is deemed to be sufficient to construct an ITS3-like detector if one would use sensors similar to MOSS, disregarding its specific faults and issues [4].
Figure 4. Functional yield of the MOSS prototype normalized per region. The pink bars illustrate the expected yield when factoring out issues, that are unique to the readout architecture of the prototype and not applicable to future chips. Taken from [4].
Functional yield of MOST was assessed by analog pulsing of all pixels of a chip. Based on data from 9 chips, it is found that on average 0.009 % of pixels per RSU are unresponsive [5].
Energy Calibration
Both sensor prototypes are able to measure the energy deposit of incident particles. Here the energy deposit is assessed using a source emitting photons with energies of 5.9 keV (Mn-K) and 6.5 keV (Mn-K). Pixels of both sensors show linearity between
Figure 5. MOSS single pixel cluster spectrum measured with photons from a source. Taken from [4].
injected charge and ToT. There is a significant pixel-to-pixel variation, which can be compensated for by performing a ToT-vs.-charge calibration pixel by pixel. Figures 5 and 6 show the obtained spectra in units of ToT respectively. In addition to the Mn-K and Mn-K peaks, in the case of MOSS the silicon fluorescence and corresponding escape peak are clearly observable. In the case of MOST hints of these peaks are identifiable. In both cases, the Mn peaks are fitted with a sum of two Gaussians. The energy resolution is calculated as FWHM/mean = , where and are mean and standard deviation of the Mn-K Gaussian fit respectively. This yields an energy resolution of for MOSS (22.5 m pitch) [4] and for MOST, where only single pixel clusters are considered in both cases.
In-Beam Testing
The MOSS sensor has been thoroughly studied in beam at CERN PS. Figure 7 shows the detection efficiency and fake-hit rate as a function of the average sensor threshold for the larger pixel pitch of 22.5 m. Green curves show a non-irradiated sample. Above 99 % efficiency and below hits/pixel/s are achieved for a range of about 80 e of thresholds. Two other samples have been irradiated to an ionizing dose of 10 kGy using x-rays or a non-ionizing dose of 1 MeV n cm using nuclear reactor neutrons respectively. As expected, the fake-hit rate increases after exposure to ionizing radiation, which likely is related to a worse performance of the front-end due to altered transistor parameters. The detection efficiency decreases after exposure to non-ionizing radiation. Furthermore, the fake-hit rate increases due to increased leakage current as well as residual radioactivity in the chip. However, an operational margin still remains after ionizing and non-ionizing radiation to ITS3 levels, as depicted by the orange and blue boxes in the background [4].
Figure 6. MOST single pixel cluster spectrum measured with photons from a Fe source.

Figure 7. Efficiency and fake-hit rate of MOSS chips obtained in a beam test at CERN PS. Taken from [4], modified.
Diode Biasing
The alternative biasing scheme of MOST (c.f. Sec. 1) has been tested by applying different combinations of AVDD and and measuring the sensor threshold. As defines the potential difference between the lower rail of the front-end and thus the collection electrode, increasing means effectively applying a reverse bias to the sensor diode. AVDD is only increased to remain a sufficient operating margin of the front-end circuitry. The resulting curve is shown in Figure 8, indicating a clear decrease of sensor threshold with increased . For future chips this biasing scheme provides an alternative as it removes the need to distribute negative voltages across the chip [5, 6].
Figure 8. Sensor threshold average in units of injection voltage as function of at different AVDD settings for MOST. Taken from [5], modified.
Power Gating
To test the power gating switches, digital and analog supply currents are measured while subsequently closing all individual switches. A smooth uniform increase is observed, as shown in Figure 9. No sudden jumps are occurring, indicating no faulty pixel groups in this chip. Also no plateaus exist, showing that all switches are closing. The effectiveness of a power switch for isolating a fault and operating the rest of the chip could not be proven as in all tested chips, not a single instance of fault has been found, that appeared only after closing the power switches. All observed faults are attributable to shorts in the power grid, for which the power switches do not provide mitigation.
Figure 9. MOST analog and digital current consumption during subsequent attaching of groups to the power grid via switches.
Conclusion
MOSS and MOST have successfully proven that building a wafer-scale stitched sensor is feasible. No problems attributed to the stitching itself have been observed. At the required radiation levels of the ITS3 ( kGy and MeV cm) MOSS remains operable with % efficiency and a fake-hit rate hits/pixel/s. The yield achieved is sufficient for constructing the detector, when disregarding issues related to the prototype readout architecture, as well as power grid related faults, which are understood and will likely be mitigated in the next sensor. In this case, a region yield of % can be achieved. The test features and operational experience of both chips provided valuable feedback for the design of the final ITS3 sensor ASIC.
Acknowledgements
This work has been sponsored by the Wolfgang Gentner Programme of the German Federal Ministry of Education and Research (grant no. 13E18CHA).
References
- [1]ALICE collaboration, Technical Design Report for the ALICE Inner Tracking System 3 - ITS3 A bent wafer-scale monolithic pixel detector, CERN-LHCC-2024-003, 2024, https://cds.cern.ch/record/2890181.
- [2]P. Vicente Leitão et al., Development of a Stitched Monolithic Pixel Sensor prototype (MOSS chip) towards the ITS3 upgrade of the ALICE Inner Tracking system, in: Journal of Instrumentation, 2023, doi:10.1088/1748-0221/18/01/C01044.
- [3]G.H. Eberwein et al., A Novel Approach for Fault Detection and Failure Analysis of CMOS Copper Metal Stacks, in: IEEE Transactions on Nuclear Science, 2026, doi:10.1109/TNS.2026.3671605.DOI
- [4]O. Abdelrahman et al., Characterisation of the first wafer-scale prototype for the ALICE ITS3 upgrade: the monolithic stitched sensor (MOSS), in: Nuclear Inst. and Methods in Physics Research, A, 2026, doi:10.1016/j.nima.2026.171297.
- [5]J. Sonneveld et al., Yield, noise and timing studies of ALICE ITS3 stitched sensor test structures: The MOST, in: Nuclear Inst. and Methods in Physics Research, A, 2025, doi:10.1016/j.nima.2025.170764.DOI
- [6]M. Selina et al., Exploring unique design features of the Monolithic Stitched Sensor with Timing (MOST): yield, powering, timing, and sensor reverse bias, in: Journal of Instrumentation, 2025, doi:10.1088/1748-0221/20/07/C07032.DOI







