2 May 2024
Job Information
- Organisation/Company
IMT Atlantique- Department
Doctoral division- Research Field
Computer science » Other
Geosciences » Other
Mathematics » Applied mathematics- Researcher Profile
First Stage Researcher (R1)- Country
France- Application Deadline
15 May 2024 - 23:00 (Europe/Paris)- Type of Contract
Temporary- Job Status
Full-time- Offer Starting Date
1 Oct 2024- Is the job funded through the EU Research Framework Programme?
Not funded by an EU programme- Is the Job related to staff position within a Research Infrastructure?
No
Offer Description
Context
In recent years, scientific computing problems have been explored from the perspective of machine learning and Artificial Intelligence (AI). The combination of AI with computational sciences has given rise to a wide spectrum of methodological questions, articulated in the framework of Scientific Machine Learning (SciML). These questions encompass various aspects, from model acceleration through AI solvers to the learning of model correction terms or sub-models in hybrid numerical modeling systems.
This project aims to extend current modeling efforts on AI-based surrogate models and hybrid modeling systems to ocean dynamics. Specifically, we aim to advance current state-of-the-art in AI-based surrogate and hybrid modeling of realistic geophysical systems [1] to ocean configurations closely resembling operational flows. The resulting models are expected to demonstrate improved forecasting and simulation performance compared to classical numerical models. We focus on regional ocean modeling configurations, including the Mediterranean Sea, the North Atlantic Ocean, and the Indian Ocean.
Scientific Objectives and Proposed Approach
The primary goal of this PhD is to advance the development of next-generation regional ocean models by building upon recent advances in surrogate modeling for geophysical systems [2] and hybrid numerical models [3]. From a methodological point of view, the specific objectives of this thesis are described below:
- Objective 1 (O1): Development and training of pure data-driven surrogate models for ocean forecasting.
- Objective 2 (O2): Development and training of hybrid Oceans Models that are based on simple physical cores.
- Objective 3 (O3): Demonstrate the performance of these models with respect to state-of-the-art models in short term forecasting and study their generalization to long term simulation applications.
In order to achieve these objectives, we propose to focus on the following questions.
Observations and Data processing
The preparation and processing of ocean data to produce training datasets is a central task in this project as it allows to benchmark, share and evaluate the proposed AI-based techniques. Following the WeatherBencsh [2] and OceanBencsh [4] projects, we focus on gap free datasets issued either from reanalysis and/or simulation products to implement supervised training strategies for the following three regions:
- Western Mediterranean configuration: The training data in the Western Mediterranean configuration will be based on the reanalysis [5].
- North Atlantic configuration: The training data in the North Atlantic configuration will be based on the high resolution NEMO simulation [6].
- Indian ocean configuration: The training data in the Indian ocean configuration will be based on high resolution simulations based on the Rregional Oocean Mmodeling Ssystem (ROMS) [7].
Development and training of pure data-driven surrogate models for ocean forecasting
Data-driven surrogates show great potential in producing accurate short-term forecasts, with a performance that is often comparable or even better than classical numerical models [2] for global forecasting applications. However, how do these models perform on regional configurations, where the state of the system exhibits different dynamical regimes? Furthermore, how do these models handle inflow/outflow conditions at the boundaries of the modeling domain?
This first objective aims to address these questions by designing pure data-driven emulators capable of handling the above challenges of regional forecasting. The performance of these models will be evaluated based on both the accuracy of the short-term forecast, but also on the relevance of the ensemble prediction of the surrogate model. The latter is very important for instance in data assimilation.
Development and training of hybrid Oceans Models that are based on simple physical cores
Beyond pure deep learning based surrogate models, we also aim at studying the impact of including a physical prior into the modeling scheme of these regional ocean models. The resultant hybrid system will be designed based on simple physical cores in order to study the impact of the prior on the training objective of the deep learning model. Regarding the form of the deep learning based sub-models, we aim at investigating both deterministic and stochastic parameterizations including physics based stochastic models defined under the Location Uncertainty (LU) principle. Overall, this objective is dedicated to the investigation of the following questions:
- Training hybrid models that are based on simple physical cores: We aim here at training hybrid models that are based on Multi-Layer Quasi-Geostrophic (MLQG) equations to fit the training data defined in O0. We use the MLQG implementation of [87] that works on arbitrary geometries. This implementation is written in Pytorch, which makes possible the definition and resolution of end-to-end learning problems.
- Deterministic vs Stochastic neural Sub-models: Deterministic sub-model corrections of physical models in Hybrid numerical modeling systems are typically formulated as correction terms that are constructed based on the resolved components of the flow. Likely, we are missing several degrees of freedom that represent the independent or generic fine-scale variability. This missing variability generates uncertainty. Taking into account and modeling this uncertainty is mandatory in applications that require probabilistic forecasting, such as data assimilation.
Development of an evaluation protocol of surrogate/hybrid models of O1 and O2 against traditional numerical modeling systems in regional case studies including the Mediterranean Sea and the Indian Ocean.
The trained models will be evaluated on both forecasting and simulation performance of the dynamics of the training data. We consider time averages statistics of variables such as Sea Surface Height (SSH), Sea Surface Temperature (SST) and Surface current. We also consider evaluating the performance of the data-driven models in simulating high resolution structures such as mesoscale eddies.
In addition to these evaluation criteria, our objective is to establish an evaluation protocol to determine the efficacy of the physical priors employed in hybrid models of O2 relative to using pure data-driven surrogates of O1. This protocol will provide insights on the advantages and limitations of incorporating physical knowledge into the modeling framework.
Thesis offer : https://www.imt-atlantique.fr/sites/default/files/recherche/Offres%20de%20th%C3%A8ses/2024/2024_Neural%20Regional%20Ocean%20Models%20(NROM).pdf
Requirements
- Research Field
- Computer science » Other
- Education Level
- Master Degree or equivalent
Skills/Qualifications
- Completion of a master's degree or equivalent in physics, mathematics, computer science, engineering, statistics, or a related field at the time of the appointment;
- Programming and/or data analysis experience;
- Interest in the application of machine learning to science and engineering problems;
- A solid record transcript record;
- Ability to work independently and as part of an interdisciplinary team;
- Ability to work in a fast-paced environment;
- Strong communication skills.
Internal Application form(s) needed
2024_Neural Regional Ocean Models (NROM).pdf
English
(175.38 KB - PDF)
Download
Additional Information
- Website for additional job details
https://www.imt-atlantique.fr/sites/default/files/recherche/Offres%20de%20th%C3…
Work Location(s)
- Number of offers available
- 1
- Company/Institute
- IMT Atlantique Bretagne - Pays de la Loire
- Country
- France
- City
- Brest
- Geofield
Where to apply
[email protected]
Contact
- City
Brest- Website
https://www.imt-atlantique.fr/en
[email protected]
[email protected]
STATUS: EXPIRED