Applying and optimizing the Exa.TrkX Pipeline on the OpenDataDetector with ACTS

Machine learning is a promising field to augment and potentially replace part of the event reconstruction of high-energy physics experiments. This is partly due to the fact that many machine-learning algorithms offer relatively easy portability to heterogeneous hardware and thus could play an important role in controlling the computing budget of future experiments. In addition, the capability of machine-learning-based approaches to tackle nonlinear problems can improve performance. Particularly, the track reconstruction problem has been addressed in the past with several machine-learning-based attempts, largely facilitated by the two highly resonant machine-learning challenges (TrackML). The Exa.TrkX project has developed a track-finding pipeline based on graph neural networks that has shown good performance when applied to the TrackML detector. We present the technical integration of the Exa.TrkX pipeline into the framework of the ACTS (A Common Tracking Software) project. We further present our efforts to apply the pipeline to the OpenDataDetector, a model of a more realistic detector that supersedes the TrackML detector. The tracking performance in this setup is compared to that of the ACTS standard track finder, the Combinatorial Kalman Filter.