Modern large-scale scientific experiments, such as those conducted at LHC (CERN) and the NICA Project, generate massive volumes of data that require efficient storage and processing. Monitoring systems in distributed computing environments accumulate valuable information about data access patterns, which can be leveraged to optimise computational workflows. One promising approach is popularity-based data management, where frequently accessed data is cached or replicated, while rarely used data is archived. However, the chaotic and irregular nature of access patterns poses challenges for traditional statistical analysis. This work explores the application of modern machine learning methods for predicting data popularity and identifying groups of datasets with similar access behaviour, enabling more efficient data caching, replication, and archiving strategies.