Neural networks for atomistic modelling - are we there yet?

As the atomistic modelling communities within chemistry, physics and material science are embracing machine learning, new methods and applications that bring machine learning to these fields are surging. In particular the neural network-based approaches, thanks to the overlap with the wider machine learning community, have experienced a notable increase in diversity.
Today, atomistic modelling approaches integrated with neural network techniques can be found in numerous studies, each employing a different flavor of these methods, reporting promising results and remaining challenges.
In this talk, we will give an overview of the current state of affairs of the field, in particular for approaches that attempt to bypass the quantum mechanical simulations. From the basic ingredients, e.g. input data, network architecture, learning targets, to dynamical factors of the neural network paradigm such as training and optimization details, we will examine the effect of the choices made on the ultimate success of the models, on the examples of systems such as carbon allotropes and molten salts. We will bring a critical eye to the current definition of success of these approaches and examine the potential of the suggested paths to make these techniques competitive.