Extreme events generated in microcavity lasers and their predictions by reservoir computing

TL;DR

This paper demonstrates the prediction of extreme events in microcavity lasers using reservoir computing, achieving advance warnings of up to 5 nanoseconds despite limited data and sampling constraints.

Contribution

It introduces a novel hybrid reservoir computing approach to predict extreme laser events, outperforming previous methods in timing accuracy.

Findings

Reservoir computing can predict extreme laser events with 5 ns lead time.

Hybrid configurations improve prediction accuracy and reliability.

Experimental and theoretical analysis confirms the method's effectiveness.

Abstract

Extreme events generated by complex systems have been intensively studied in many fields due to their great impact on scientific research and our daily lives. However, their prediction is still a challenge in spite of the tremendous progress that model-free machine learning has brought to the field. We experimentally generate, and theoretically model, extreme events in a current-modulated, single-mode microcavity laser operating on orthogonal polarizations, where their strongly differing thresholds -- due to cavity birefringence -- give rise to giant light pulses initiated by spontaneous emission. Applying reservoir-computing techniques, we identify in advance the emergence of an extreme event from a time series, in spite of coarse sampling and limited sample length. Performance is optimized through new hybrid configurations that we introduce in this paper. Advance warning times can reach 5ns, i.e. approximately ten times the rise time of the individual extreme event.