Emulating high-resolution ocean dynamics from the available large-scale, multimodal and multiscale observation and simulation databases – EMOCEAN

During the last decade, the ocean community witnessed worldwide the launch of over 30 new ocean-related satellite missions. Plans for new satellites, to improve the spatial-temporal sampling, are already laid well into the foreseeable future, and today, we are already talking Petabytes of data to download, analyze, transform into accessible information. Increasing computer power and understandings of relevant physical processes are also rapidly evolving, and contribute to advances in model accuracy and resolution refinements. For instance, the Japanese Earth Simulator, now creates global complex coupled models of ocean and atmosphere with unprecedented details, shedding light on fundamental mechanisms to explore high resolution observations.

Within the next decade, past, actual and future satellite Earth Observation missions, extended in situ networks and super-computer simulations shall thus continue to pave this new era to understand the Earth system as a whole, to serve both research and operational interests. In this context, this project aims at challenging the current thinking in ocean data assimilation to anticipate the new generation of data-intensive applications.,We propose to address the development of big-data-related learning-based tools and methods. With a view to actually exploiting the potential of massive simulations and data streams, new tools shall be developed to complement existing capabilities and provide observation-driven emulations and/or decompositions of high-resolution upper ocean dynamics.

Based on our preliminary developments, this project shall dwell on recent theoretical/methodological advances, especially in the fields of physical oceanography, space oceanography, signal and image processing and machine learning, as well as on big-data-oriented technologies. As foreseen, our project aims at designing and operating optimal solutions for fast and robust high-resolution "EMulations of the OCEAN” (EMOCEAN) from massive heterogeneous, multimodal and multiscale observation and simulation data.

From the multidisciplinary expertise in our consortium, the methodological core of this proposal will be the development of novel tools and methods to:
• learn from massive observation and/or simulation data new multi-sensor/multi-modal/multi-scale representations of the upper ocean dynamics (Objective O1). Supported by a theoretical analysis of ocean dynamics, these representations will typically consist of a collection of deterministic and stochastic transfer functions characterizing the interactions between different geophysical features and/or different space-time scales;
• exploit the novel representations to emulate at high-resolution the upper ocean dynamics from combined low-resolution operational model ouputs and/or incomplete satellite snapshots (Objective 02). The learnt representations will serve as novel data-driven basis functions to project partial operational observations and/or predictions, to explore and recognize the underlying dynamical forcings in space and time, and emulate realistic high-resolution predictions of the targeted geophysical fields.
• implement these strategies within a high-performance architecture to address the massive data pouring from space and simulations (Objective O3). This objective comprises both considering a big-data-compliant architecture, especially the learning step (Obj. O1), and accelerated instantaneous implementation for the end-users (Obj. O2).
The proposed methodologies will be specified, implemented and evaluated within two different case-studies, regarded as proofs-of-concept:
• The emulation of high-resolution daily sea surface composite fields from multi-sensor data, building on available multi-sensor satellite observations and model predictions (CS1);
• The short-term emulation of particle dispersal scenarii at the ocean surface building on new composite fields (CS2).