Data-driven prediction of reversal of large-scale circulation in turbulent convection

TL;DR

This study demonstrates that reservoir computing can predict the reversal of large-scale circulation in turbulent convection using only local sidewall measurements, with successful short-term predictions and insights into the challenges of long-term forecasting.

Contribution

It introduces a novel application of reservoir computing to predict LSC reversals in turbulent convection using sparse, local measurements, highlighting feasibility for experiments and industry.

Findings

Short-term prediction of LSC reversal is successful using local measurements.

Long-term prediction often fails due to statistical independence of reversals.

Prediction accuracy decreases after the reversal in long quasi-stable states.

Abstract

Large-scale circulation (LSC) quasi-stably emerges in the turbulent Rayleigh-B\'{e}nard convection, and intermittently reverses its rotational direction in two-dimensional turbulent convection. In this paper, direct numerical simulations of the intermittent reversals of the LSC in a two-dimensional square domain are performed, and the time series of the total angular momentum indicating the rotational direction of the LSC is predicted by reservoir computing whose input consists of the shear rates and temperatures at six locations on the sidewalls. The total angular momentum in the simulation after times shorter than half the typical duration of the quasi-stable states is successfully reproduced by the locally-measurable quantities on the sidewalls because the secondary rolls accompanied by the boundary flow characterize the reversal of the LSC. The successful prediction by such sparse input derived from local measurements on the sidewalls demonstrates that the reservoir computing prediction of the reversal is feasible in laboratory experiments and industrial applications. On the other hand, long-term prediction often provides the total angular momentum opposite in sign to the one in the simulations in the late parts of long quasi-stable states. The statistical independence of each reversal implies that the prediction after the reversal is difficult or even impossible, and the training data in the late part in the long quasi-stable state, which rarely appears, is contaminated by the statistically-independent angular momentum in the subsequent quasi-stable state.