Research

Can we extend the forecast range of AI weather models?

Today's AI weather models forecast well a few days out, but skill drops sharply at subseasonal-to-seasonal (S2S) lead times of two to six weeks — exactly the range most useful for planning around monsoons, heatwaves, and other slow-moving hazards. We're exploring architectures borrowed from other domains, including retrieval-augmented generation and token-based latent representations drawn from language modeling, to push S2S skill further than current approaches allow.

Working on this: Pinak Mandal, Utpal Anand, Adhin AS, Lakshya Chouhan, Vaibhav Dekhawat

Representative work: Inductive biases in deep learning models for weather prediction

Can we make AI weather models more reliable for extremes?

Most AI weather models are benchmarked on average forecast skill, but the events that matter most — cyclones, extreme rainfall — are rare by definition. We evaluate how well current models, used as released, actually capture these events, from cyclogenesis over the Bay of Bengal to South Asian monsoon extremes.

Working on this: Ratul Tarafder, Subhajit Biswas

How can we leverage teleconnections?

Weather in one region is often linked to conditions thousands of kilometers away — the Indian monsoon's connection to ENSO, or synchronized rainfall extremes between North India and the Sahel. We study these teleconnections directly, and how weather-state networks and intraseasonal oscillation indices can be forecast and used to anticipate regional impacts.

Working on this: Hrithuparna Bharat, Divyansh Yecho

Representative work: Intraseasonal synchronization of extreme rainfall between North India and the Sahel, Teleconnection patterns of different El Niño types revealed by climate network curvature

How can we make AI weather models more efficient?

State-of-the-art AI weather models are large and expensive to run, putting them out of reach for many research groups and forecasters — particularly across the Global South. We work on pruning and physics-constrained architectures that cut computational cost without sacrificing forecast accuracy, including coupling fine-scale climate predictions to downstream socioeconomic models.

Working on this: Abhishek Menon, Ankit Kumar

How can we make AI weather models more reliable?

Beyond raw accuracy, do these models know when they don't know? We work on calibrating AI weather model uncertainty with statistical guarantees, and on understanding how forecast skill degrades as the underlying data distribution shifts over time — and how to adapt models to compensate.

Working on this: Dheeraj Kumar Gehlot, Nishidh Singh