Fig. 5: Reduced uncertainty in El Niño-Southern Oscillation (ENSO) sea surface temperature (SST) amplitude projections using artificial neural networks (ANNs).
From: Projection of ENSO using observation-informed deep learning
a Projected ENSO SST amplitudes under the high-emission scenario. In the left panel, the blue shading represents the inter-model spread (the one standard deviation range) across 11 Coupled Model Intercomparison Project (CMIP) models, with the thick blue line representing the multi-model mean. The red shading represents the spread across 11 outputs from ANNobs, which is the ultimate model that incorporates all ANNs and describes the real-world ENSO response to tropical Pacific warming patterns. The thick red line represents the ensemble mean. The black shading represents the range of three observational datasets. The right panel shows the probability density functions (fitted to Gaussian distributions) for projected ENSO SST amplitudes. b Comparison of ENSO SST amplitudes between ANNobs and CMIP models. BEST6 and REM5 represent the CMIP models with realistic and unrealistic ENSO physics, respectively. Amplitudes are normalized by their 1994–2023 values. ANNobs projects a reversal in ENSO SST amplitude trends around 2050 (dashed line; P1 and P2 denote periods of increasing and decreasing amplitudes, respectively), aligning with the BEST6 models. In contrast, the REM5 models project a much later reversal. c ANN-based dependence of ENSO SST amplitudes on the east-minus-west SST gradient and the SST mean state in the eastern equatorial Pacific (EP). The scatter plots represent the averaged projection over the 11 ANNobs outputs (corresponding to the black line in panel (b), but without normalization), and shading represents the fit of ANNobs outputs to a binary linear function. The gray dashed line marks the year 2050, showing that reversal occurs when the trajectory is tangent to the contour lines. Deep learning not only reduces uncertainty in ENSO projections through ANN-based constraints but also provides a straightforward insight into the non-unidirectional evolution of ENSO SST amplitudes in the 21st century.
