“The Work of Thought”: How Machine Learning Is Transforming Our Understanding of the Earth System

Redacción HC
04/10/2025

Machine learning (ML) is reshaping the way scientists study our planet’s climate and weather. A new academic opinion piece published in PLOS Climate (Oldham-Dorrington J, Quinting J, Sobolowski S, 2025) explores how this technology could usher in a new era of Earth system science. Far beyond simply improving forecasts, the authors argue that ML can drive a deeper theoretical understanding of how the planet works—if researchers integrate it carefully with traditional physical models.

From Forecasting to Fundamental Insight

In recent years, ML-based prediction models have outperformed traditional physics-based approaches in high-resolution weather forecasting. For example, the European Centre for Medium-Range Weather Forecasts (ECMWF) developed the AIFS system, which uses ML to deliver cutting-edge predictions. Similar advances have emerged in modeling rainfall at kilometer scales and emulating complex processes such as ocean temperature evolution.

These successes raise a crucial question: will the surge in data-driven methods weaken the causal, theory-based understanding that has historically underpinned climate science? According to the authors, the answer could be the opposite. They argue that machine learning, far from undermining theory, may catalyze a “work of thought” that strengthens our conceptual grasp of Earth’s dynamic systems.

Building on Open Data and Collaborative Science

One of the key enablers of this revolution is the availability of massive open datasets like ERA5 and WeatherBench. Such resources allow computer scientists—even those with limited training in Earth sciences—to build state-of-the-art forecasting models. This democratization of data is accelerating innovation and lowering the barriers to entry for high-quality climate modeling.

However, the authors caution that these models must be rigorously tested for “generalizability”—their ability to make accurate predictions beyond the conditions on which they were trained. Standardized tests could include reproducing unobserved hurricanes, chaotic error growth, or dynamic instabilities. Without such tests, ML models risk hidden biases and compensating errors that could mislead decision-makers.

A Positive Feedback Loop Between Data and Theory

The paper envisions a productive feedback cycle: ML models can uncover patterns that inspire new physical hypotheses, while physical theory provides criteria to evaluate and guide machine learning. In this hybrid framework, physics-based models remain essential. They will generate synthetic data, provide boundary conditions for downscaling, and form the backbone of hybrid modeling strategies.

This shift allows scientists to focus on fundamental questions—slow climate variability, paleoclimate reconstruction, and counterfactual climate scenarios—while letting ML handle the computationally intensive tasks of prediction and pattern recognition.

Implications for Policy and Society

If machine learning models can be made robust and physically grounded, they could transform how societies prepare for extreme weather and long-term climate change. More accurate and affordable forecasts would improve early warning systems for hurricanes, floods, and droughts, helping farmers, urban planners, and emergency responders.

The authors emphasize that interdisciplinary collaboration is vital. Climate scientists and computer scientists must work together to design standardized generalization tests and to maintain the development of physical models as a complement—not a competitor—to machine learning.

Conclusion: A New Era for Climate Science

The “machine learning revolution,” as described in the PLOS Climate article, is not merely a technical upgrade. It represents a paradigm shift in how humanity can understand the Earth system. By blending data-driven insights with rigorous physical theory, scientists can move beyond better forecasts to deeper conceptual understanding.

This “work of thought,” as the authors put it, has the potential to elevate Earth system science into a new era of creativity and discovery. The challenge now is to ensure that machine learning models are not only powerful but also explainable and reliable across the uncharted climates of the future.

Topics of interest

Climate