Samuele Santacà

Samuele Santacà works on data-driven modelling of cyber-physical systems, with an emphasis on industrial systems whose behaviour is too heterogeneous or complex to describe accurately through a single first-principles model. His research explores how operational data can be turned into dynamic virtual representations that retain the temporal behaviour and cross-channel relationships of the physical system.

His thesis work develops a framework based on transformer architectures for learning these virtual models directly from multivariate time-series data. Self-attention is used to capture long-range temporal dependencies and correlations among signals that may be missed by simpler recurrent or convolutional approaches. The resulting models are intended for digital-twin applications including simulation, predictive analysis and decision support without requiring complete a priori knowledge of the plant dynamics.

A second part of the work investigates real-time anomaly detection at the edge. This includes model compression, quantisation and inference optimisation for constrained devices, with the objective of preserving useful predictive accuracy while meeting latency, memory and power limits. The broader research direction combines machine learning with the ESD group’s work on deterministic industrial simulation and software validation, aiming at virtual models that can contribute to more adaptive and resilient cyber-physical production systems.