Deep learning approaches to top FCNC couplings to photons at the LHC

We investigate the sensitivity of the LHC to flavour-changing neutral current interactions involving the top quark and a photon using a model-independent effective field theory framework, focusing on two complementary processes: single top production via $qg\to t\gamma$ and the rare decay $t\to q\gamma$ in top pair events. To enhance signal discrimination, we employ a range of deep learning classifiers, including multi-layer perceptrons, graph attention networks and transformers, and compare them against a traditional cut-based analysis. Our results demonstrate that attention-based architectures, in particular transformer networks, significantly outperform other strategies, yielding up to a factor of five improvement in the expected exclusion limits. In particular, we show that at the high-luminosity LHC, rare top branching ratios can be probed down to values as low as ${10}^{-6}$. Our results thus highlight the significant potential of attention-based architectures for improving the sensitivity to new physics signatures in top quark processes at colliders.