Department of
Data Science

Bedartha Goswami focuses on finding principled machine learning methods to derive low-dimensional representations of the climate and on developing data-driven models to predict weather patterns from long-term to seasonal scales. He received his Masters degree from IISER Pune in 2011 followed by a PhD at the Potsdam Institute for Climate Impact Research in 2015. Before joining IISER Pune in October 2024, he led an independent research group on machine learning and climate science at the University of Tübingen.

Research

Machine Learning in Weather and Climate 

Low dimensional representations (LDRs) of the climate system: Through this work, Dr. Bedartha Goswami estimates sparse representations of climate datasets using principal components, non-negative matrix factorisation, or correlation-based networks. LDRs allow evaluating climate model outputs more meaningfully and also possibly reveal new features of climate variability that are not observable at first sight from the data.

Early warning signals (EWSs) of extreme rainfall: LDRs often allow us to discover precursors of extreme events which could be developed into simple empirical models for early warnings. These EWSs would essentially be of the form: “Monitor observable X in region A, and when it exceeds some threshold and when a set of other background conditions are met, we have a high likelihood of an rainfall extreme event N days later in region B.”

Inductive biases of deep learning weather models: While deep learning weather models are now at par with numerical weather prediction models, their inner workings are opaque and their limitations unclear. Dr. Goswami investigates the biases of deep learning weather models so that they can develop more robust and interpretable variants. Adversarial robustness and memorisation-generalization trade-offs are a central focus in finding inductive biases.

Foundation model for South Asian weather: Foundation models are pre-trained via self-supervised learning to learn useful data features that help to perform well at ‘downstream’ tasks. The group is building a foundation model for South Asian weather, which will be able to predict extreme rainfall, cyclone tracks, and heat waves.

Selected Publications

Schlör J, Strnad F, Capotondi A, Goswami B. Contribution of El Niño Southern Oscillation (ENSO) diversity to low‐frequency changes in ENSO variance. Geophysical Research Letters. 2024 Jul 28;51(14):e2024GL109179.

Haas M, Goswami B, von Luxburg U. Pitfalls of climate network construction—a statistical perspective. Journal of Climate. 2023 May 15;36(10):3321-42.

Strnad FM, Schlör J, Geen R, Boers N, Goswami B. Propagation pathways of Indo-Pacific rainfall extremes are modulated by Pacific sea surface temperatures. Nature Communications. 2023 Sep 15;14(1):5708.

Strnad FM, Schlör J, Fröhlich C, Goswami B. Teleconnection patterns of different El Niño types revealed by climate network curvature. Geophysical Research Letters. 2022 Sep 16;49(17):e2022GL098571.

Boers N, Goswami B, Rheinwalt A, Bookhagen B, Hoskins B, Kurths J. Complex networks reveal global pattern of extreme-rainfall teleconnections. Nature. 2019 Feb 21;566(7744):373-7.