Include telescope pointing
Milestone ID: 2055
Previous studies [9] showed that γ-PhysNet is very sensitive to the difference of telescope pointing direction between the MC dataset used for training and the actual telescope pointing direction during observations. This task aims at establishing a strategy to overcome this issue. Candidate solutions include training multiple networks in parallel on a grid of MC data and automatically applying the closest one to the data at inference time, or developing a large network including the telescope pointing direction as variable input to be trained on MC data from the whole sky. These strategies involve different levels of complexity in engineering and their application in production will depend on the required and available resources (memory, training time, etc...). An important part of task 4.1.1 will be to test these different options and determine the best trade-off.