CIO Insider

CIOInsider India Magazine


IBM and NASA's New AI Weather Model Could Unravel Climate Issues


IBM and NASA are joining hands for a new Artificial Intelligence project. The goal of IBM’s climate initiatives and NASA earth science is to create a multimodal basic AI model for weather and climate prediction that can be scaled too many downstream tasks with relatively few GPUs and climate forecasting. IBM’s AI experts will work closely with climate scientists and other NASA experts to test and validate the AI weather forecasting models in seven applications, including 10-14 day weather forecasting and things like climate resilience, dust storms, and air turbulence.

Central to these initiatives is IBM's geospatial foundations model, developed in collaboration with NASA. Trained on vast amounts of geospatial data such as satellite imagery, these models aim to derive environmental insights and solutions related to climate change. Unlike traditional AI models, they use extensive climate-related data to accelerate the analysis of various environmental aspects affected by climate change.

Climate Technologies
Alessandro Curioni, IBM Fellow and Vice President of Accelerated Discovery, explains the potential impact of these technologies on understanding and addressing climate-related events. Basic AI models using geospatial data can be a game-changer, allowing us to better understand climate-related events with unprecedented speed and efficiency.

The new AI model is data from NASA satellites for geospatial intelligence, and according to IBM, it is the largest geospatial model on the open-source AI Hugging Face platform. So far, the model has been used to monitor and visualize tree planting and cultivation activities in water tower areas, such as forested landscapes that retain water in Kenya. The aim is to plant more trees and solve water shortage problems. The model is also used to analyze urban heat islands in the United Arab Emirates.

The Renaissance in Weather Forecasting
Weather forecasting has improved dramatically in recent decades. Today's six-day forecast is as accurate as a five-day forecast 10 years ago. Hurricane tracks can be predicted more accurately three days ahead than 24 hours and 40 years ago. This remarkable achievement is due to two things: decades of progress in atmospheric and ocean science and concurrent advances in high-performance computing. Modern weather models base their predictions on massive computer simulations that take time and energy to run. That's because they take into account both physical equations and weather observations, from wind and air pressure to temperature and precipitation.

Basic models have several advantages that come from their ability to process and analyze raw data of many types, allowing them to create a broad representation of data that can be generalized to many scenarios. It is an important capability in a field such as climate, where conditions are constantly changing in time and space, and many downstream applications exist outside of predictions.

Inferring Atmospheric Dynamics from the Data
To be basic, the base model cannot be a simple pony. It should be able to perform many tasks and ideally be trained on many types of data. This is particularly important in weather and climate prediction, as many physical processes can often only be observed at certain time frames and spatial scales. For example, the cyclical El Niño weather pattern occurs over many months and over half the globe, while tornado initiation can take minutes and originate from sub-meter-scale processes.

IBM's geospatial model powers a digital platform for tracking tree planting activities, helping local reforestation efforts, and measuring the impact of carbon sequestration

The sensors provide a continuous, highly localized record of changing temperatures, wind, and pressure. Satellite images, on the other hand, capture changes in the environment at longer intervals and at a lower resolution. IBM and NASA's proposed baseline model will initially be trained on the MERRA-2 dataset, a combination of high-quality observations and estimates of past weather over the past 40 years. Observational data from fixed weather stations, floating weather balloons, and satellites orbiting the planet will be added later. IBM and NASA are currently experimenting with model architectures and techniques to integrate these different temporal and spatial scales into a single multimodal model.

Analyzing Urban Heat Islands in UAE
IBM and Mohamed Bin Zayed University of Artificial Intelligence (MBZUAI) are working together to map urban heat islands in Abu Dhabi using a fine-tuned version of IBM's geospatial foundations model. The goal is to understand the influence of the local landscape on temperature anomalies. Initial results indicate a reduction in heat island effects and provide insights for future urban design strategies.

Reforestation and Water Sustainability in Kenya
In partnership with Kenya's Special Envoy for Climate Change, Ali Mohamed, IBM is supporting the National Tree Growing and Restoration Campaign. The initiative aims to plant 15 billion trees by 2032, especially in critical areas of water towers affected by deforestation. IBM's geospatial model powers a digital platform for tracking tree planting activities, helping local reforestation efforts, and measuring the impact of carbon sequestration.

Elevating Climate Resiliency in the UK
In partnership with the Science and Technology Facilities Council (STFC) and Royal Haskoning DHV, IBM is developing AI-driven tools for climate risk assessment in the UK. These tools will initially focus on assessing the effects of weather on air traffic. In addition, the TreesAI research project aims to map areas suitable for planting trees to mitigate surface water flooding and offers a digital planning platform for urban developers.

Current AI models often miss extreme events. This tendency is a known problem with AI models that are trained to ignore outliers. Loss functions minimize the chance of making big mistakes, but then they can also miss extreme events. Methods correcting this tendency have been implemented in smaller models. A challenge from IBM and NASA will extend this work to large base models.

Another problem is climate change itself. The past isn't always a great predictor of the future, especially when the climate is warming as fast as it is today. For example, a hurricane in 2024 may have higher wind speeds than a hurricane in 1933. As a result, forecasters may not see it if their models are based solely on historical data. However, AI allows models to be continuously updated as circumstances evolve and new data becomes available.

Current Issue
ARETE: Pioneering Cyber Risk Solutions & Transforming The Future Of Cybersecurity