The path from scientific to operational flare forecasting: a deep learning approach

In our view, machine/deep learning for flare forecasting is still more a promise for future scenarios than the reference framework for current operational facilities. This delay from the application of AI methods in research settings to their use for real-time forecasting is probably due to the persistence of technical open issues involving, by instance, the optimization strategy of the training phase, the quantitative assessment of the prediction performances, the reduction of the computational burden. This talk proposes a video-based deep learning approach to flare forecasting in which the optimization of the network’s parameters is realized by means of a probabilistic score-oriented loss function, the training procedure accounts for the part of the solar cycle progression when the prediction is requested, and the prediction performances are assessed by means of value-weighted skill scores that give greater importance to the values of the prediction than to its quality. The talk will also show the operational potentialities of this approach and discuss how feature selection may reduce the information redundancy, thus increasing the computational efficiency.