by Denisse Sciamarella, Marisol Osman and Juan Ruiz

translated from https://www.clarin.com/revista-n/podemos-predecir-cambio-climatico-inteligencia-artificial-irrumpe-metereologia_0_chsBY1dKhg.html

22 April 2024 marked the 300th anniversary of philosopher Immanuel Kant’s birth. In the Critique of Pure Reason, Kant asks what we can know. Artificial intelligence, the term we use today to refer to any system that implements mechanisms close to those of human reason, revives Kant’s question in a new way.

Algorithms can execute automatic processing tasks that outperform human capabilities in speed, achieving surprising results through training often called “machine learning”. Machines “learn” by compilation and recombination, by work prepotency —as Arlt would say— but without “understanding” anything. How necessary is “understanding” to extract knowledge from observations? Here is a line of rupture in what we mean by knowing, which multiplies into many new questions in each disciplinary field.

Machine learning techniques in marine and atmospheric sciences represent a unique opportunity to extract knowledge from large volumes of available data. For example, ocean and inland water levels have never been measured so accurately. Automatic processing of observations and simulations allows the development of weather forecasting models whose performance is comparable to that of the physical process-based models traditionally used in the world’s weather services. How far can we go with machine learning-based models without knowing anything about the physical processes, and what applications can these models have beyond forecasting a few days in advance?

 Unlike weather forecasting, the definition of “climate” cannot be given without understanding the long-term behaviour of the Earth system. In projecting future climate change, knowledge about the physics of the climate system embedded in physical-mathematical simulations will continue to play a central role. In other words, studying climate change is possible only by identifying the mechanisms that explain the variability in time evolution.

A branch of mathematics called topology is also in full development. Topology is used to study the “shape” of data and, very recently, improve the performance of machine-learning algorithms. On the other hand, the chaos theory also recurs to topology, but in a much more fundamental way, using the “shape” behind the data to understand the mechanisms that lead to a given type of dynamics.

These developments make it possible to imagine a virtuous link between the use of artificial intelligence —with no understanding of the intervening phenomena, as in the case of short-term forecasting— and the use of chaos topology to access the underlying processes that provide the basis that enables addressing the problem of climate change and its impacts. Returning to the Kantian question, we can risk a conclusion: the breakthrough represented by artificial intelligence once again calls upon the capacity of human beings to understand, as creators of knowledge.