Data-driven Global Ocean Modeling for Seasonal to Decadal Prediction

TL;DR

ORCA-DL (Oceanic Reliable foreCAst via Deep Learning), a data-driven three-dimensional ocean model for seasonal to decadal prediction of global ocean dynamics, accurately simulates the three-dimensional structure of global ocean dynamics with high physical consistency and outperforms state-of-the-art numerical models in capturing extreme events.

Abstract

Accurate ocean dynamics modeling is crucial for enhancing understanding of ocean circulation, predicting climate variability, and tackling challenges posed by climate change. Despite improvements in traditional numerical models, predicting global ocean variability over multi-year scales remains challenging. Here, we propose ORCA-DL (Oceanic Reliable foreCAst via Deep Learning), the first data-driven 3D ocean model for seasonal to decadal prediction of global ocean circulation. ORCA-DL accurately simulates three-dimensional ocean dynamics and outperforms state-of-the-art dynamical models in capturing extreme events, including El Niño-Southern Oscillation and upper ocean heatwaves. This demonstrates the high potential of data-driven models for efficient and accurate global ocean forecasting. Moreover, ORCA-DL stably emulates ocean dynamics at decadal timescales, demonstrating its potential even for skillful decadal predictions and climate projections.

Figures (27)

• Figure 1: Climatological mean states of 1985-2018. (A to B) The mean states of sea surface temperature (SST; shading) and sea surface currents (SSC; vector) in the June-July-August (JJA) and December-January-February (DJF) seasons based on the predictions from ORCA-DL at a lead time of 12 months. (C to D) The bias of SST mean-states in the ORCA-DL's predictions at 12 months lead relative to GODAS in the JJA and DJF seasons, respectively. (E to F) The scatter plots of the SST and sea surface salinity (SSS) climatology in the JJA season based on GODAS and ORCA-DL's predictions at 12 months lead, respectively. The dashed contour lines denote potential density with a reference pressure of 0 dbar ($\sigma_0$ density, $\rm kg/m^3$). (G to H) As in (E) to (F), but for climatology in DJF season.
• Figure 2: Distributions of temporal Anomaly Correlation Coefficient (ACC) skills of ORCA-DL for temperature and salinity prediction. (A to C) The ACC skills of ORCA-DL for sea surface temperature (SST) predictions at lead time of 6, 12, and 18 months, respectively. The black contour line is the dividing line where the prediction skill is equal to 0.5. (D to F), (G to I) and (J to L) As in (A) to (C) but for the prediction skill of temperature at a depth of 300m ($\rm T_{300}$), SSS, and salinity at a depth of 300m ($\rm S_{300}$), respectively.
• Figure 3: Skills of ORCA-DL in predicting ENSO. (A) The DJF Niño 3.4 index based on GODAS and predictions of ORCA-DL at lead times of 3, 6, 12, and 24 months, respectively. The x-axis represents the winter season of each year, i.e., 2015 represents the period from December 2015 to February 2016. (B) Correlation skills of Niño 3.4 index prediction as a function of the lead month. The testing period is 1985-2018. (C to D) The 20° C isotherm depth (d20) anomalies averaged over latitude (5° S-5° N) in the Pacific basin based on GODAS and ORCA-DL prediction at a lead time of 6 months respectively. The x-axis and y-axis denote the time and longitude, respectively. (E to G) The depth-longitude distribution of ACC skills of temperature prediction in the equatorial region at lead time of 6, 12, and 18 months, respectively.
• Figure 4: Forecast skills of upper ocean marine heatwaves. (A) The distribution of the frequency of the occurrence of upper ocean marine heatwave events based on GODAS from 2010 to 2018. (B to H) The regionally average SEDI skill as a function of the lead time for seven selected regions with a relatively high frequency of heatwave occurrences. The skills are calculated based on MHW events for the period between 1985 and 2018. The higher SEDI denotes the better skill.
• Figure 5: Decadal prediction of ORCA-DL. (A) The global mean SST based on GODAS and ORCA-DL's prediction with initial field of December 1989, respectively. The solid line indicates the ensemble mean results while the shadow indicates the spread of ORCA-DL's different members. (B) The global annual mean of SST anomalies based on GODAS and ORCA-DL's predictions at lead times of 1, 3, and 5 years. The time range is selected as 1990-2019 to unify the forecasts for different lead times. (C to E) Correlation skills of the Pacific Decadal Oscillation (PDO), Interdecadal Pacific Oscillation (IPO), and Atlantic Multidecadal Oscillation (AMO) index forecast as a function of lead time, respectively. The testing period is also 1990-2019. The skills shown in the bar chart are calculated using annual mean (1y and 2y) and multi-year mean (2-5y etc.) data, with the x-axis representing lead years. In contrast, the skills depicted by the red line with square markers are based on monthly data, with the x-axis representing lead months (2m, 6m, etc.).