Artificial intelligence and meteorology: where AI outperperforms humans and where it still falls short
Artificial intelligence can describe intricate natural processes with a precision that often surpasses human capability. Yet it cannot invent new equations and test them against the established laws of physics. This view comes from a leading figure in the field of weather science who oversees Russia’s Hydrometeorological Center. The assessment highlights both the power and the current limitations of neural networks in scientific inquiry.
The expert notes several domains where neural networks and machine learning demonstrate clear advantages over human analysis, especially in situations marked by ambiguity. In these moments, when data patterns are intricate and the rules governing the system are not yet fully defined, artificial intelligence tends to excel at inference and pattern recognition.
The discussion centers on processes in which fundamental laws do not already exist in a ready form or are still being formulated. Subgrid-scale processes, which occur at scales smaller than the practical resolution of weather models, serve as a prime example where neural networks can fill gaps by learning from data rather than relying on explicitly written equations.
In such contexts, the neural network can produce useful insights by recognizing relationships in vast data sets and making predictions even when the governing mathematics is incomplete or evolving. This capability makes AI a strong partner in weather forecasting, especially for capturing nuanced patterns that are hard to codify into traditional equations.
However, the same lines of reasoning illuminate a fundamental boundary. Current artificial intelligence does not possess an intrinsic understanding of the core laws of physics. It does not conduct theoretical research in the way scientists in meteorology and physics do, by formulating hypotheses, deriving equations, and testing them against physical principles and empirical data. The core equations that researchers derive, such as those related to energy conservation and momentum conservation, reflect universal physical laws that neural networks do not inherently produce or validate. They emerge from human insight and rigorous mathematical derivation rather than from machine learning alone.
The prevailing view is that progress will come from synergy. By combining the strengths of artificial intelligence with the expertise of meteorologists and physicists, it is possible to achieve a collaborative workflow where models learn from data while researchers guide and constrain them with physical laws and experimental design. This intersection promises more accurate forecasts and a deeper understanding of atmospheric processes, while keeping the integrity of fundamental physics at the core of scientific practice.
In a related note, an American tech company recently explored an unusual application of artificial intelligence by experimenting with identifying products described in poems and ballads. This example underscores the broad potential of AI to interpret diverse kinds of textual information and extract meaningful patterns, a capability that can inform natural language processing and data interpretation in many fields, including meteorology.