Description
Spectral diffusion is a key limitation for the integration of solid-state quantum emitters into nanophotonic platforms. In diamond nanopillars, nitrogen-vacancy (NV) centers exhibit enhanced spectral instability due to surface-induced electric-field fluctuations. It is an open question whether stochastic or correlated effects dominate the noise and to what degree.
In this work, we employ a recurrent neural network with a Long Short-Term Memory (LSTM) architecture, a feed-forward neural network (FFNN), and a Transformer to model the temporal evolution of the zero-phonon line (ZPL) frequency and to determine if predictions beyond coincidence can be made. All models achieve only limited predictive performance, reflecting the intrinsic randomness of the system. However, qualitative differences emerge: recurrent architectures yield more stable predictions, attention-based models are sensitive to temporal scale, and feed-forward approaches exhibit greater variability.
The results highlight both the challenges of modeling spectral diffusion in nanostructures and the importance of architecture choice when applying machine learning to quantum systems.