Research Projects

Since 2023, three different weather models based on deep learning architectures have been published in high impact journals, viz. Pangu Weather (Nature, [1]), GraphCast (Science, [2]), and Neural GCM (Nature, [3]). This was the result of efforts by both university- and industry-based research groups worldwide, such as Google, NVIDIA, and Huawei, to leverage the recent success of deep learning models for weather prediction.These deep learning weather models (DLWPs) offer a drastically different paradigm for weather prediction where, instead of using physical equations to model atmospheric dynamics, we train a deep neural network with terabytes of weather data to detect patterns that are useful in predicting variables of interest. Moreover, DLWPs offer the advantage that, once trained, they are much faster and computationally cheaper than traditional weather models at issuing predictions. This is the reason why DLWPs are being adopted by meteorological agencies around the world. For instance, the ECMWF has recently released AIFS: the first operational AI-based weather forecasting setup [4]. The success of the DLWP models have thus far been assessed primarily on global benchmarks such as WeatherBench2 [5], and their performance on regional weather systems, particularly in the Global South, is unknown. Barring GraphCast, most of the DLWP models do not model precipitation (Neural GCM outputs effective precipitation, i.e. precipitation-minus-evaporation, as a derived variable). In South Asia, where rainfall plays a predominant role in society and environment, it is pertinent that we develop a DLWP that predicts precipitation reliably. And a key component to predicting precipitation well is to have a “generative” DLWP model which will not suffer from “blurring” of weather states typically seen in DLWPs over longer forecast horizons (greater than 10 days). To develop a DLWP for South Asian and India, we thus need to: 1. Assess the performance of existing SOTA DLWPs over South Asia and fine tune them to improve performance 2. Develop a generative long-range DLWP forecast model that does not suffer from blurring. Such a model will be usable by stakeholders in various parts of society at low computational cost to generate local weather forecasts attuned to their needs. REFERENCES [1] Bi, Kaifeng, et al. "Accurate medium-range global weather forecasting with 3D neural networks." Nature 619.7970 (2023): 533-538. [2] Lam, Remi, et al. "Learning skillful medium-range global weather forecasting." Science 382.6677 (2023): 1416-1421. [3] Kochkov, Dmitrii, et al. "Neural general circulation models for weather and climate." Nature 632.8027 (2024): 1060-1066. [4] Lang, Simon, et al. "AIFS-ECMWF's data-driven forecasting system." arXiv preprint arXiv:2406.01465 (2024). [5] Rasp, Stephan, et al. "WeatherBench 2: A benchmark for the next generation of data‐driven global weather models." Journal of Advances in Modeling Earth Systems 16.6 (2024): e2023MS004019