BSC to collaborate in an ESA project for the provision of Generative Artificial Intelligence in space domain

31 October 2019

BSC will provide its emerging technologies and HPC infrastructure for conducting the activities of this project.

The Barcelona Supercomputing Center (BSC) will investigate the use of Generative Artificial Intelligence for generating and augmenting synthetic datasets for remote sensing applications in space domain. BSC together with its Italian partners AIKO S.r.l and the National Research Council - Institute of Atmospheric Sciences and Climate are the members of the consortium of the DeepLIM project, funded by the European Space Agency (ESA)

The Computer Architecture and Operating System (CAOS) group is in charge of this project at BSC and will provide the centre’s emerging technologies and HPC infrastructure including POWER9 processors and NVIDIA V100 GPUs for conducting the activities of this project.

Training deep learning models, due to the nature of the problem, is very computation-intensive. As a consequence, deep learning workloads have been shown to be an excellent fit for accelerators such as GPUs included in the MareNostrum P9 Cluster of Emerging Technologies.

BSC will help with the performance optimizations of deep learning libraries which will result in a faster and more energy-efficient training and inference. By performing a thorough analysis of the main deep learning libraries, BSC researchers will identify the performance bottlenecks and most time-consuming functions and by optimizing them for specific architectures (e.g., GPU or CPU), both training and inference processes can be done in a more efficient way.

Deep learning approaches are the primary solution in many domains and recently, they are being used in critical domains such as space.

DeepLIM project has two main objectives. First, understanding, developing and exploiting the use of Generative Artificial Intelligence to improve and augment datasets acquired via observation campaigns, or generated by computationally-intensive models. Second, improving the state-of-the-art models that are in use to perform Inversion Modelling with the use of Deep Learning algorithms. In particular, an improvement in data acquisition campaign for simulation and algorithm development and cost reduction due to the lower amount of real data required for inversion models training is expected.