Distributed training and scalability for the particle clustering method UCluster

In recent years, machine-learning methods have become increasingly important for the experiments at Large Hadron Collider (LHC). They are utilised in everything from trigger systems to reconstruction and data analysis. The UCluster method is a general model providing unsupervised clustering of particle physics data, that can be easily modified provide solutions variety different decision problems. current paper, we improve on by adding option training scalable distributed fashion, thereby extending its utility learn arbitrarily large sets. combines graph-based neural network called ABCnet with step, using combined loss function phase. original code publicly available TensorFlow v1.14 has previously been trained single GPU. It shows accuracy 81% when applied problem multi-class classification simulated jet events. Our implementation adds functionality utilising Horovod framework, which necessitated migration v2. Together parquet files splitting up between compute nodes, makes any amount input something will essential use real LHC We find well suited training, time decreasing direct relation number GPU's used. However, further improvements more exhaustive possibly hyper-parameter search required order achieve reported method.