HSF participant pages for GSoC 2026

AI-Accelerated Signal Reconstruction for the ATLAS Tile Calorimeter at the HL-LHC

XBastille

The ATLAS Tile Calorimeter turns digitised pulses into energy measurements forty million times a second, inside an FPGA. At the High-Luminosity LHC the algorithm doing that today falls apart, because the noise it was designed for is no longer the noise that exists. I spent the summer replacing it with a 91-parameter neural network, getting that network into fixed point, placing and routing it onto the real device, and merging a bit-exact emulator of it into ATLAS Athena. Here is how that went.

Apache Arrow interface for PODIO

Arnav Dham

This project aimed to implement a new Apache Arrow backend for PODIO (Plain-Old-Data I/O), a C++ library designed for high-performance data modeling in particle physics. Over the summer, I added an in-memory Apache Arrow backend(Along with parquet writing and reading capability) to enable language-independent, columnar data access for flat and nested structures, improving throughput and enabling modern heterogeneous computing workflows.

Automated Software Performance Monitoring for the ATLAS Experiment

Douglas Lindsay

Prior to this project, software regression testing at ATLAS was predominantly performed manually via time-consuming and error-prone inspection of metrics on the ATLAS Performance Monitoring Board, which provided no easy way to identify potential root causes ofregressions.In collaboration with the Software Performance Optimisation Team (SPOT), I modified the SPOT orchestration scripts to incorporate an automated anomaly detection pipeline that autonomously detects anomalies, issues alerts to SPOT, and even attempts to identify the root cause itself. A range of statistical and ML techniques were explored for identifying regressions and ultimately three regression-detection algorithms were developed. I also upgraded the Performance Monitoring Board to provide at-a-glance visibility of recent anomalies.

Debuggable Installations For Spack Packages

Sebastian Paucar

A new machine- and build-system-agnostic debugging standard for Spack packages is implemented. DWARF-referenced in/out-of-tree and generated source code, along with split debug symbols, are preserved on demand alongside automated GDB configurations. An OCI buildcache infrastructure is built for debug info distribution (fetch/push).

EKO Oxidation

Akshat Rai

This project aimed to modernize EKO (Evolution Kernel Operators) by bridging its Python codebase with a Rust backend. Over the summer, I implemented a robust cross-language interoperability layer, developed two dedicated interface crates for C-ABI and Python bindings, backed by automated CI/CD workflows that deploy directly to PyPI and GitHub Releases.

Fine grained storage for the DUNE experiment

Ahmed Idani

DUNE is building a new data processing framework, Phlex, whose I/O layer FORM lets each pipeline stage write its own data product into its own container without waiting for anyone else. That makes writing cheap, but it moves the cost onto the reader: someone who needs several products of the same event can follow only one of them in its natural order and has to read the rest out of order. I spent the summer building a benchmark on ROOT’s RNTuple to find out what that actually costs. Short version: you pay in decompression, not in disk, and the size of the bill is set by how many events share a page. At one event per page the read order is free. At four thousand, the same bytes cost 48 times as much.

Generative-AI Assisted Testing of Complex Spack Packages

Vaishnavi Mishra

Traditional Continuous Integration (CI) test suites typically validate only “leading-edge” software configurations (the newest package versions, default variants, and the latest compilers). This project aims to build a Spack extension that leverages Large Language Models (LLMs) and a self-adaptive feedback loop to autonomously explore and test high-risk, off-leading-edge configurations, discovering undeclared incompatibilities and compiler regressions that traditional pipelines miss.

Integration of CMS Combine with FCCAnalyses

Soumyadip Niyogi

This summer, as a Google Summer of Code contributor with the HEP Software Foundation at CERN, I am bridging the gap between FCCAnalyses and CMS Combine. My project aims to build a native Python interface that automates the generation of datacards and RooFit workspaces directly from RDataFrame histograms, streamlining the path from simulated events to physics results.

Full list of projects in 2026

… will be published in Nov 2026.