Speakers
Description
As the CMS pixel detector approaches the end of Run 3, it sustains increasing radiation damage that significantly impacts tracking performance. The resulting degradation in charge collection efficiency leads to "broken clusters" (pixels missing within a cluster) and "shorter clusters" (missing end pixels). These effects degrade track resolution and introduce substantial hit position bias, particularly in high-η regions where tracks traverse the silicon sensors at shallow angles. Figure 1 illustrates this effect: as the readout threshold increases relative to the diminished charge profile of an irradiated sensor, clusters progressively lose pixels and eventually split into fragments.
Figure 1: Charge loss after irradiation in a CMS pixel module. Top: schematic of a flipped module showing charge collection in the local y (global −z) and local x (global φ) directions. Bottom: cluster profile in z showing how increasing thresholds progressively break an irradiated cluster into fragments.
We present a novel, Alpaka-based [1] pixel "digi-morphing" algorithm designed for the CMS High-Level Trigger (HLT) to recombine these fragmented clusters. The algorithm operates directly at the individual pixel hit level using an image-morphing "closing" technique. It is a morphological operation consisting of a dilation step followed by an erosion step, each employing convolution kernels. The dilation expands cluster boundaries to bridge small gaps between fragments, while the subsequent erosion restores the cluster shape, retaining only the newly connected pixels. Figure 2 illustrates this procedure: a 3×3 kernel is applied to first dilate the broken cluster, filling in the gap, and then erode the result to recover the original cluster extent with the gap bridged.
Figure 2: Illustration of the morphological closing operation applied to a broken pixel cluster. A 3×3 dilation kernel expands the cluster boundaries (middle), and a subsequent erosion step (right) restores the shape while retaining the recovered pixels that bridge the gap.
While an offline version of the algorithm existed previously, the HLT deployment required a complete rewrite targeting the Alpaka portability framework to enable parallel execution on both CPUs and GPUs. The implementation supports regional morphing, allowing selective application to the most radiation-damaged detector regions which are currently the outermost module rings of the two first layers of the pixel barrel. Performance benchmarks demonstrate a negligible impact on HLT throughput with morphing enabled. Physics validation indicates a roughly 6.2% increase in average cluster size, confirming the successful recombination of fragmented hits without introducing spurious clusters in unaffected regions.
The current implementation employs two highly optimized packed 1-bit-per-pixel status buffers stored as a 1D array of 32-bit words, drastically reducing the memory footprint while maintaining algorithmic equivalence. A binary search over a sorted list of module identifiers is used to efficiently determine whether a given module should undergo morphing.
In this contribution, an explanation of the issue and algorithm will be presented, followed by a description of its implementation, and the study of its impact on performance and timing of local-reconstruction, as well as higher level track parameters.
The algorithm has been integrated into the CMS software framework (CMSSW) [2,3] and subsequently refined to fix the ongoing race conditions in edge case scenarios causing HLT crashes [4,5].
References
[1] E. Zenker, B. Worpitz, R. Widera, A. Huebl, G. Juckeland, A. Knüpfer, W.E. Nagel, and M. Bussmann, "Alpaka — an abstraction library for parallel kernel acceleration," 2016 IEEE International Parallel and Distributed Processing Symposium Workshops (IPDPSW), pp. 631–640 (2016)
[2] Digi Morphing for HLT, CMSSW PR #48734 (2025)
[3] Backport: Digi Morphing for HLT, CMSSW PR #48832 (2025)
[4] HLT digi morphing fix, CMSSW PR #49991 (2026)
[5] Backport: HLT digi morphing fix, CMSSW PR #50045 (2026)
The authors acknowledge the CMS Tracker DPG group and Andrea Bocci for contributions to the optimized implementation.