Heliophysics is a vast topic that covers the study of the interior of the Sun, solar physics, the interplanetary medium, solar wind-magnetosphere interactions, the dynamics of magnetosphere and its coupling with the ionosphere-thermosphere system, and even the space weather effects on space and ground-based infrastructures.
Like all other disciplines, heliophysics is not immune to the machine learning revolution that is currently taking place in science. Indeed, nowadays many recognize that the techniques and tools developed by computer scientists (often to tackle completely different applications), combined with big data and easy access to accelerated computing, are becoming the fourth pillar of scientific discovery. More importantly it is now abundantly clear that the data-driven approaches investigated by means of machine learning techniques will become more and more mainstream and, therefore, are here to stay.
The heliophysics community is now facing the challenge of overcoming the barrier of technical skills posed by machine learning that are not generally mastered by the typical scientist. We therefore need to fully appreciate and critically understand what is within reach in a few years and what could be achieved in a decade.
Heliophysics is a very interdisciplinary topic, where the focus of a research effort can range from basic understanding of physical phenomena to devising operational space weather capability able to forecast specific events. The number of ways that machine learning can be generally leveraged in the Earth and space sciences was showcased in Bortnik and Camporeale (2021).
The new cross-journal special collection “Machine Learning in Heliophysics” follows the 2nd edition of a community-driven conference of the same name that was successfully held in Boulder, Colorado in March 2022. The conference program and many oral and poster presentations are still available to download. The major takeaway of the conference was that the field is rapidly moving from an exploratory phase where machine learning techniques were often unsuccessfully attempted on problems that were not well designed for such approaches to a more mature phase that yields a much higher success rate and compelling results.
Among all the possible uses of machine learning techniques in heliophysics, we emphasize the few that we believe have the potential of true scientific breakthrough within the next few years:
- Reduced order modeling, or acceleration/emulation of computationally expensive physics-based models
- Physics-informed machine learning, where physics-based constraints are encoded in a machine learning architecture. This is often referred to as a ‘grey-box’ approach that bridges purely data-driven (black-box) and physics-based (white-box) approaches, with the goal of leveraging the advantages of both approaches and limiting their weaknesses
- Data-driven discovery of new physical laws and/or new parameterization of physical quantities
Other topics relevant to the special collection include inverse estimation of physical parameters, automatic event identification, feature detection and tracking, time series analysis of dynamical systems, combination of physics-based models with machine learning techniques, surrogate models and uncertainty quantification.
This is a joint special collection between Space Weather, JGR: Space Physics, Geophysical Research Letters, and Earth and Space Science. Manuscripts can be submitted to any of these journals, depending on their fit with the journal’s scope and requirements.
—Enrico Camporeale (email@example.com; 0000-0002-7862-6383), University of Colorado, USA; Veronique Delouille (0000-0001-5307-8045), Royal Observatory of Belgium, Belgium; Thomas Berger (0000-0002-4989-475X), University of Colorado, USA; and Sophie Murray (0000-0002-9378-5315), Dublin Institute for Advanced Studies, Ireland