AI has been receiving significant negative attention recently, and has been portrayed as embodying the worst aspects of technology. Large Language Models (LLMs), for instance, require vast amounts of data, computing power, and human ingenuity to train, costing hundreds of billions of dollars. The benefits of these models seem limited to more efficient search queries, task management, and software usability. Meanwhile, they've sparked numerous concerns: implicit biases, questionable judgment, plagiarism in publishing, and social isolation. The energy consumption required for LLM training and inference also raises serious questions about future greenhouse gas emissions.
Despite this negative perception, there's a less publicized aspect of AI: its application in scientific research and machine learning. When combined with subject-specific knowledge, AI has enabled significant progress in numerous scientific fields.
Consider numerical weather prediction, an area of meteorology that uses atmospheric physics to predict the weather. This field examines atmospheric fluid dynamics, electromagnetic radiation (sunshine and heat) movement through the atmospehre, and microphysical processes involving water droplets and aerosols. Over the past 75 years, scientists have developed complex equations that enable numerical weather prediction, forming the basis for the weather forecasts we rely on daily.
In fact, the full set of equations that describe the atmosphere cannot be solved on pencil and paper, but requires the use of supercomputers. Until recently, there were only a dozen major weather forecasting centers around the world, like the National Weather Service, which had the necessary high performance computing resources to produce weather forecasts over different regions of the globe. There were also a handful of private companies that produced more specialized forecasts than those produced by these national and regional centers.
The rise of scientific AI and machine learning has revolutionized weather prediction. These technologies enable the creation of statistical models trained on observed weather data, surpassing traditional physics-only approaches in both accuracy and computational efficiency. By exploring various statistical model architectures and training data methods, AI weather models now outperform even the most advanced numerical weather prediction systems.
Similar breakthroughs have occurred in neuroscience, materials science, drug discovery, and genetics, where AI has accelerated scientific progress. Unlike LLMs and other generative AI used for content creation, applying AI to scientific disciplines doesn't necessarily increase energy consumption.
Climate science is another field where AI models have the potential to be disruptive. Earth System Models (ESMs) simulate the complex interactions between ocean, atmosphere, land, ice, and biosphere that determine our planet's environmental conditions. These models solve thousands of physics equations using over a million lines of code, requiring millions of core hours on supercomputers.
Unlike AI-based weather models, we lack the petabytes of observational data to create pure AI-based Earth system models, but that doesn’t mean AI can’t play a valuable role in improving ESMs. Machine learning models can better capture the complexity of real-world clouds and aerosols compared to idealized physics equations. Incorporating these ML models into larger physics-based ESMs combines the strengths of both approaches. Alternatively, hybrid AI models use physics equations alongside observational data, requiring less training data than purely statistical models while being over 1000 times more computationally efficient than conventional ESMs. This approach saves millions of core hours of computational resources and hundreds of simulation hours.
At Planette, we leverage AI and machine learning to operationalize environmental forecasting for longer time frames – from 1 month to 5 years into the future. By blending AI with subject matter expertise, our models achieve greater efficiency and accuracy than the best sub-seasonal and seasonal forecasting systems worldwide, including those of NOAA and ECMWF. We're proud to be part of the scientific revolution made possible by AI and machine learning.