How Big Tech Companies are Tackling Climate Change

An increasing number of tech companies are making their mark on the climate sector, leveraging the transformative power of advanced AI/ML models. Recently, Huawei announced its latest AI model Pangu Weather. The model boasts a higher precision when compared with traditional numerical weather forecast methods. 

The model utilises deep learning techniques along with 43 years of historical data with a prediction speed considered 10,000 times faster than traditional methods. According to China Meteorological Administration, Pangu Weather had successfully forecasted the path of the recent Typhoon Mawar with the precision of five days prior to its alteration in the waters near Taiwan’s Eastern islands. 

Google Research developed a deep learning model MetNet-2 that focuses on utlising AI for weather and climate-related issues. The model can predict precipitation with remarkable precision, providing forecasts at a spatial resolution as fine as 1km and a time resolution of 2 minutes for a duration of up to 12 hours. 

Microsoft and DeepMind have also built their AI models with ClimaX and Graphcast, respectively. 

Meanwhile, NVIDIA is working on the Earth 2 model in an effort to collaborate with climate researchers and policymakers. Through the utilisation of Modulus and FourCastNet (ML model that imitates the dynamics of global weather patterns predicting extremes with unprecedented speed and accuracy), NVIDIA achieved significant improvements in the weather trajectory generation. They were able to produce 21-day weather trajectories for 1,000 ensemble members in a fraction of the time it previously took for a single trajectory. 

IBM has also been in the climate space for a while now. In 2016, IBM acquired The Weather Company, which was a subsidiary of the Weather Channel. With IBM’s Watson framework, the model uses AI to combine information from over 100 weather forecast models worldwide. 

Startup Zeus AI, started by former NASA scientists, leverages vast volumes of data from government satellites, including information about atmospheric winds, water vapours, temperature changes, and cloud cover that impact weather affect global weather patterns.  Some of the other companies solving weather prediction problems include Tomorrow.io, Atmo.io, Jua.ai and Zeus.ai. 

Limitations 

Implementing ML models for weather prediction comes with its limitations. Limitation in training the data is where the problem lies. Rare and extreme weather conditions poses a challenge in training and testing. Furthermore, data availability is another problem. In NWP models, where satellite data is used, the missing values are interpolated. However, by using such interpolated data for AI models, the phenomenon of concept drift and built-in biases can arise. 

According to a paper on machine learning for weather and climate modelling, the authors suggest that neural networks may require explicit instruction on relationships between certain variables. Short-to-medium datasets are insufficient to enable a model to understand long-term variations such as El Nino or any form of climate change. 

The New-Age AI Forecasting 

Weather forecasters primarily rely on numerical weather prediction (NWP) models. These models use data from weather stations, weather balloons, and satellites to understand the current state of the atmosphere. By solving equations related to air movement, these models can predict most weather patterns accurately. However, through this method, smaller weather events such as localised thunderstorms, or predicting which side of a town will experience heavy rain during a thunderstorm is challenging to predict. The method also involves expensive computational power, and that is where AI models have an advantage. 

Forecasting methods using ML models can analyze a large dataset of past weather maps to understand typical weather patterns, as opposed to traditional NWP models thats solve complex physical equations to arrive at patterns. By training on historical data, AI models will then use current weather information to make future predictions. However, this method also lacks the ability to forecast localised weather conditions. 

The machine learning systems use significantly less computing power when compared to traditional prediction methods. NVIDIA’s Earth-2 model reduces energy consumption by approx 1000 times. 

India Meterological Department (IMD) has started using AI as an experimental phase to improve nowcast and short-range weather forecasting which range from three hours to seven days. 

AI models can provide accurate and tailored forecasts for specific user needs by looking at fine-grained details in the data that traditional methods may overlook, however, with the challenges that exist, AI models will most likely not be able to replace traditional methods as of now.