Work Package 3
Evaluation of The Data and Model Prediction
Participant : Thales, CNRS, DLR, DWD, TUD
WP leader : Thales
Start month : 01 – End month : 46
Main contact: firstname.lastname@example.org
Design AI algorithms to extract and combine relevant information such as contrails localization, dynamics and persistency from data collected as part of WP1 and/or available from public sources.
Build hybrid-AI algorithms to identify contrails and distinguish them from cirrus from data such as satellite images, ground based camera images, and/or LIDAR data
Provide robustness, encoding symmetries to allow training AI algorithms with fewer data (frugal learning), which is of particular interest in our context.
Description of work :
T3.1 Data preparation
This task includes the cleaning of the available data, the linkage of their metadata when several sources of information are used, and their labelling for supervised learning algorithms. Further relevant pre-processing steps such as renormalization, PCA etc., will also be covered in this task.
T3.2 AI algorithms design
In this task, we will work on building adequate algorithms for the considered tasks by embedding geometrical and/or physical priors where appropriate. For instance, contrails identification from satellite images in spherical geometry may be handled with Equivariant Neural Networks to avoid distortion effects due to planar projections, while their dynamics prediction may benefit from AI-hybridization with winds data and clouds motion dynamics.
T3.3 AI algorithms implementation and training
This task includes training the algorithms of T3.2 on the data prepared in T3.1. This will typically be done on GPU, by leveraging the implementation of the algorithm in usual frameworks such as TensorFlow or PyTorch.
T3.4 AI algorithms validation
AI algorithms outputs can be subject to significant uncertainty, and a thorough validation process is therefore required to infer an adequate validity domain for the trained algorithms. To do so, we will consider a statistical approach and compute corresponding confidence metrics, which will then be used to increase the reliability of the outputs and to take adequate mitigating action where appropriate.
T3.5 AI-based comparison of prediction models outputs with observations
This task evaluates the prediction algorithms developed in WP2 by comparing their outputs with the empirical knowledge extracted from the observations with the AI algorithms developed and validated as part of T3.1 to T3.4.