< Back to previous page
Project
Deep learning combined with physical modelling for weather, climate and geophysics applications (FEDTWIN2)
Current weather and climate systems produce increasingly heterogeneous and voluminous, data, e.g. multimodal satellite data but also data harvested from unstructured sources s.a. smartphones and private weather stations. The joint analysis and assimilation of these data into a numerical weather prediction model poses huge challenges. Given RMIB’s ambition to adopt an open data policy and to strengthen its role as data hub for weather, climate and geophysical data, research on modern data analytics tools is essential to create added value to RMIB’s research output and services.
RMIB’s in-depth knowhow on physics-based nonlinear numerical modelling and prediction should be extended by data-driven machine learning (ML), such as deep learning (DL), that has proven its capability to aggregate and process heterogeneous data in many applications. In conjunction with physical models, these data-driven methods are able to increase performance in various ways by capturing the intractable nonlinearities, hidden in the data. The research of VUB-ETRO on new DL paradigms will benefit from RMIB’s well-documented huge datasets and from the know-how on physical modelling to design domain-specific DL methodologies.
DL is an evolution of Neural Networks which has recently come to maturity. It is a generic set of tools which transforms input data into representative and compact output data of lower dimensionality compared to the input, where the transformation model is not fully imposed a-priori through physical modelling, but learned from training datasets instead. DL outperforms traditional methods in many applications.
DEEP’s objective is to establish a long-term research program where DL is applied in the Application Domains (ADs) of interest to RMIB, namely, Weather Forecast, Climate Monitoring and Space Weather. Currently the methods used in these ADs are mostly based on a-priori physical modelling. The overarching research challenge will be to enhance existing physical modelling through combining them with DL methods. This requires a far going integration of research expertise in both domains, through a dedicated joint long-term research position, as detailed in the proposal.
DL techniques that will be studied include: Data Transformation, Fusion and Prediction, Combined Physical and DL Modelling - in particular Residual Learning, Interpretable Deep Learning, Hybrid Methods - and Process Emulation.
As a first two year application project, a DL system for the observation and seamless nowcasting/forecasting of the amount and type of precipitation based on satellite and radar observations, as well as NWP models will be developed. Since the previous submission of DEEP (score 5.5/6), substantial new results have already been obtained in nowcasting of the amount and type of precipitation. Application projects will be defined every four years in the three ADs to refine and extend the generic research toolset and to tune it to various ADs and their requirements. Another objective is the improvement of the quality of RMIB’s services and the creation of new ones. The joint research will be supported by VUB-ETRO’s knowledge/technology transfer strategy. It will also be international from the start: RMIB will become partner of the joint laboratory involving VUB, Duke University, UCLondon and UGent.
RMIB’s in-depth knowhow on physics-based nonlinear numerical modelling and prediction should be extended by data-driven machine learning (ML), such as deep learning (DL), that has proven its capability to aggregate and process heterogeneous data in many applications. In conjunction with physical models, these data-driven methods are able to increase performance in various ways by capturing the intractable nonlinearities, hidden in the data. The research of VUB-ETRO on new DL paradigms will benefit from RMIB’s well-documented huge datasets and from the know-how on physical modelling to design domain-specific DL methodologies.
DL is an evolution of Neural Networks which has recently come to maturity. It is a generic set of tools which transforms input data into representative and compact output data of lower dimensionality compared to the input, where the transformation model is not fully imposed a-priori through physical modelling, but learned from training datasets instead. DL outperforms traditional methods in many applications.
DEEP’s objective is to establish a long-term research program where DL is applied in the Application Domains (ADs) of interest to RMIB, namely, Weather Forecast, Climate Monitoring and Space Weather. Currently the methods used in these ADs are mostly based on a-priori physical modelling. The overarching research challenge will be to enhance existing physical modelling through combining them with DL methods. This requires a far going integration of research expertise in both domains, through a dedicated joint long-term research position, as detailed in the proposal.
DL techniques that will be studied include: Data Transformation, Fusion and Prediction, Combined Physical and DL Modelling - in particular Residual Learning, Interpretable Deep Learning, Hybrid Methods - and Process Emulation.
As a first two year application project, a DL system for the observation and seamless nowcasting/forecasting of the amount and type of precipitation based on satellite and radar observations, as well as NWP models will be developed. Since the previous submission of DEEP (score 5.5/6), substantial new results have already been obtained in nowcasting of the amount and type of precipitation. Application projects will be defined every four years in the three ADs to refine and extend the generic research toolset and to tune it to various ADs and their requirements. Another objective is the improvement of the quality of RMIB’s services and the creation of new ones. The joint research will be supported by VUB-ETRO’s knowledge/technology transfer strategy. It will also be international from the start: RMIB will become partner of the joint laboratory involving VUB, Duke University, UCLondon and UGent.
Date:1 Oct 2021 → Today
Keywords:Earth Sciences
Disciplines:Other earth sciences not elsewhere classified, Meteorology, Machine learning and decision making