AGU is excited about the launch of the Journal of Geophysical Research: Machine Learning and Computation, which is now officially open for submissions. Our newest journal is dedicated to research that explores data-driven and computational methodologies based on statistical analysis, machine learning, artificial intelligence and mathematical models, with the aim of advancing knowledge in the Earth and space sciences.
JGR: Machine Learning and Computation will accept papers proposing novel results in the broad fields of solar and space physics, geophysical fluids and planetary environments, Earth surface and interiors, and biogeosciences.
Founding Editor-in-Chief Enrico Camporeale is a research scientist at Queen Mary University of London and at CIRES, University of Colorado, Boulder. Matthew Giampoala, AGU’s vice president of publications, spoke with Dr. Camporeale about how the journal came to light and what lies ahead.
AGU: What’s your research background and what inspired you to propose this new journal?
Camporeale: My educational background is in space plasma physics, though I have crossed different discipline boundaries over the years, such as moving from studying fundamental processes in space physics to the more applied field of space weather. Similarly, I moved from more standard computational methods in plasma physics to the emergent field of scientific machine learning.
I have been inspired by seeing many scientists embracing the view that machine learning will become, in time, the fourth pillar of scientific discovery, hence following a path that in the last few decades has been walked by computational physics. Although one could argue that such a process will occur naturally (and possibly inevitably), the process can definitely be accelerated by creating a dedicated venue for publications of works that do not necessarily fit in the current publication scenario.
AGU: What’s the difference between ChatGPT and the type of research that you and other Earth and space science researchers do?
Camporeale: ChatGPT and other generative AI tools have had the immense merit of exposing AI to the general public. Though extremely powerful and impressive, tools like ChatGPT merely generate text or images that are plausible, but not necessarily accurate (nor true), by learning through a number of examples that a human would take thousands of years to read. This inevitably raises some important concerns about AI safety and ethics. Whether generative AI is applicable to the process of scientific discovery is still an open question. As an example, it is well known that ChatGPT can generate plausible but completely made-up references when asked a question about scientific literature!
I would say that the (non-generative) tools used in other branches of AI (e.g., machine learning) are, at least in the short-term, more appropriate to scientific problems, being a powerful tool to explore and understand the large amount of data at our disposal and possibly a more robust way of discovering new physics from data.
AGU: What are the most exciting recent discoveries where AI has helped lead the way?
Camporeale: I am particularly excited by the ability of machine learning to discover new algorithms. In the era of computational science, every hypothesis or theory needs to be codified in the form of a computationally efficient algorithm to be verified or falsified. In turn, ideas that are not computable inevitably translate to ‘‘not falsifiable,” and hence non-scientific. Devising algorithms and improving upon existing ones is one of the most creative processes in which the human species excels. Yet AI has recently proven to be able to outsmart decades (or even millennia) or human wisdom. Two examples that come to mind are AI beating a world class champion at the game of Go and improving upon the 50-year-old open problem of matrix multiplications.
Other breakthroughs in the field of physical sciences are happening, for instance, in the field of numerical weather prediction (NWP). That is considered by many one of the major achievements of our science-based technological society. Indeed, NWP combines a very accurate understanding of the underlying physics of the atmosphere with modern advances in high-performance computing and data assimilation. Several groups around the world have now devised ML-based emulators or accelerators of physics-based NWP simulations, with similar or better accuracy. One of the most interesting advantages of using ML for prediction is that the computational time is reduced by orders of magnitude. In turn, the availability of such methods (also in terms of open-source software) is going to revolutionize our world since, unlike physics-based models, they almost require no expertise to run and can be executed on commodity hardware.
Finally, I believe that what we are witnessing is that AI is making an impact in fields that are easily accessible and understandable by the general public (chatbots being the primary example) but it is also rapidly moving towards more niche applications in almost all fields of science. ML is currently at the center of a virtuous cycle of rapid innovation, and I am very positive that it will soon start playing a major role in scientific discovery across all areas of geoscience.
AGU: Why does AGU need a separate journal with this scope? Aren’t these techniques published in other AGU journals?
Camporeale: As ML techniques are becoming more and more widespread, we have certainly seen an increase of papers that use such techniques across all AGU journals. However, those papers typically focus on the open scientific questions in their respective domain and use ML as an off-the-shelf tool to perform their analysis. A distinctive category is one where novel ML techniques need to be developed (often by way of tailoring or improving existing techniques) in order to tackle a specific scientific question. In other words, the development of the technique becomes pivotal to the advancement of the science. Such papers often had a hard time being properly reviewed and accepted in the current AGU editorial landscape. This is why we believed that a dedicated journal was needed. Moreover, new ML and computational techniques applied to a particular domain could be transferred to adjacent domains, hence it is important to not disperse literature across many journals but to offer to the community a single, easily identifiable focus. That is the ambition of JGR: Machine Learning and Computation.
AGU: How has your experience with organizing conferences and special collections shaped your perspective on the need for a new journal?
Camporeale: As an organizer and attendee of conferences, sessions and special collections, I have realized that in my field of ML for space physics and space weather applications it is very hard to stay updated and have a vision of how the field is moving, once again because of the lack of a unique editorial venue that collects similar papers. Talking with colleagues, I have recognized that the problem was widespread across all areas of geoscience and therefore there was enough scope and request from many AGU communities for such a new journal. Since the journal was launched on the AGU website and social media, I have had a great number of people reaching out to me (or meeting at conferences) manifesting how the journal was both timely and needed, which makes me very optimistic about its future impact in the community.
AGU: AGU has quite a process for developing and approving journal launches. What did you learn from this process? Was any of the feedback particularly helpful?
Camporeale: The one thing that I learned is that, as with many things in life, you cannot possibly make everybody happy! The process to develop and approve the journal launch was long and thorough but definitely a good exercise in how to compromise and take feedback in a decision-making process. The meetings with the publication committee were particularly informative, as that is a well-sized group of well-informed members of different AGU communities. One important decision was about how to name the new journal, as we were looking for keywords that would still be relevant in many years in the future. I am glad that the majority of people involved in the final decision were finally persuaded that Machine Learning is a keyword that is going to stay in science for a long time.
AGU: How can people get involved? Do you still need more editorial board members?
Camporeale: Yes, we will shortly appoint an initial editorial board. However, at this point it is hard to predict the volume of manuscript that we will receive, so we will keep track of all applications received for editors and associate editors and possibly appoint more as the journal grows. We have received an overwhelming response, once again a testament of the need for this new journal.
I encourage anybody that is unclear whether their work would fit in the journal’s scope to get in touch with me or with the deputy Editor-in-Chief Raffaele Marino and to not hesitate to ask any questions.
AGU: What sessions are you most excited about at the upcoming AGU23 meeting?
Camporeale: The number of sessions that contain the words “Machine Learning” is now overwhelming and distributed across all sections (more than 350 sessions). I am keeping an eye particularly on the ones that encourage works in the burgeoning field of physics-informed or physics-guided ML, which in my opinion, encompasses the many techniques that will lead to the next wave of scientific discoveries across many geoscience disciplines.
Aside from scientific sessions, there will be a number of wide-reaching town halls on topics centered on AI that I think will be extremely interesting. For instance, the topic of Ethical AI/ML in the Earth, Space, and Environmental Sciences (TH13G) is extremely important. Beside sessions and town halls, AGU has always excelled in bringing extraordinary keynote speakers and AGU23 will be no exception. Among all the exceptional speakers, I am really looking forward to hearing Prof. Stephen Pyne talking about the role of fire for humanity.