GPU-Accelerated Triggerless Data Acquisition Pipeline for Machine Learning-Based Anomaly Detection

Data acquisition systems in high-energy physics experiments have traditionally relied on hardware and software triggers to reduce the data rate by selecting events of interest in real time. While effective, this approach introduces an inherent bias due to the online selection criteria, potentially discarding rare or unexpected phenomena. To mitigate this limitation, triggerless data acquisition systems have been proposed, in which detector data are continuously streamed and processed without prior selection. Within this context, data acquisition pipelines can incorporate machine learning–based anomaly detection algorithms to identify deviations from nominal detector behavior or to flag potentially interesting physics signatures in an unbiased manner. However, the continuous nature and high throughput of triggerless systems impose stringent requirements on computational performance and latency. This thesis work involves the design and implementation of a triggerless data acquisition pipeline in which all processing stages (ranging from data reconstruction and preprocessing to anomaly detection) are executed on GPU-accelerated architectures. The anomaly detection stage is implemented by means of machine learning models, fully integrated with the pipeline. The performance of the system will be evaluated in terms of throughput, latency, and scalability under realistic operating conditions. The pipeline will be validated using data collected from a cosmic muon detector within the CMS collaboration at CERN. Scalability studies will be conducted to assess the suitability of the proposed approach for deployment in high-rate, large-scale particle physics experiments.