Data Driven Simulation of Semiconductor Optical Amplifiers by Means of bi-LSTM and Transformer Machine Learning Models

The recent growth of cloud-computing exponentially increased the datacenter traffic demand. New datacenter applications are gaining traction where signal loss is critical such as optical switching implemented in photonic integrated circuits. Semiconductor Optical Amplifiers (SOAs) find a new important role as they can compensate signal loss across the optical link [1] but due to their non-linear nature, their response can be modelled with complex numerical travelling wave models (TWM) [2] which lead to long computational time per run. Data-driven machine learning methods can address the limitations of numerical simulations in modelling complex physical systems, reducing the convergence time and potentially increasing the replication-prediction accuracy. In this work, we propose what is to our knowledge the first data driven modelling approach based on bidirectional long-short term memory (bi-LSTM) and self-attention transformer networks to fully simulate SOAs in the temporal domain. We train the models to reproduce the power dynamics of a measured SOA device, predict the bit-error-ratio (BER) accurately in PAM -4 amplification and enable reliable simulation across multiple operating conditions almost 45 times faster compared to a TWM [2].