Decoding complexity: how machine learning is redefining scientific discovery is also the topic of a review article written as an international collaborative effort between members of SeRC (KI, KTH, SU), the Norwegian Meteorological Institute (MET), Helmholtz Center Munich/Technical University of Munich (TUM), University of Washington (UW), and the Massachusetts Institute of Technology (MIT).
The pre-print is available on arxiv.org. The article provides a broad overview about the application of AI methods in various scientific fields. It suggests that the choice of ML method depends on the availability of fundamental equations for a given problem. The authors present a gradient of equation knowledge and present examples of AI applications sorted along this axis.
The article also draws attention on the amount of data that is produced by modern experiments compared to historical examples and emphasizes that computational speed can be a breakthrough in itself if it makes the processing of vast data sets possible.
Various different techniques can be summarized under the umbrella term artificial intelligence. Specific examples from several fields of study are connected to the specific ML methods used to enable discovery, speed up numerical methods, or gain knowledge about mechanisms of observed phenomena, among others: deep reinforcement learning, generative adversarial networks, large language models, reduced order models.
The article presents a table which locates these method types on the gradient of availability of fundamental equations in the field of study, e.g.: physics constrained learning is useful when fundamental equations are at least partially applicable to the problem at hand.
This review is a good preparation for the upcoming annual meeting.
