Large Quantum Networks

Overview:

The classical approach to modeling the dynamics of physical systems relies on intricate physical models that capitalize on robust prior knowledge of underlying phenomena and numerical simulations demanding extensive computations. However, the proliferation of vast datasets sourced from satellite imagery or local sensors has spurred the exploration of data-intensive methodologies. The aim is to complement traditional physical-based simulations or create emulators for cost-effective approximations of dynamics.

The exploration of neural networks has resulted in a growing body of work, as evident by the contributions highlighted by [1]. These methods have demonstrated efficiency in learning the intricate, nonlinear, multi-scale structures inherent in geospatial fields. Applications in satellite data, spatial products, and numerical models, as well as various tasks such as clustering, inversion, and downscaling have shown promise.

In practice, these models are often optimized with co-located spatially and temporally in-situ measurements, assuming a negligible effective lag between the two geophysical fields or considering uniform spatial resolution. However, many of these contributions are confined to simple problems or variations of existing methodologies. Successful deployments in climate studies are infrequent. Attempts to directly apply methods from other fields, like computer vision, typically fall short of surpassing traditional approaches. The majority of relevant works tend to prioritize enhancing global performance, often overlooking geophysical interpretation, regional performance evaluation, and the model’s behavior in unseen extreme events.

The incorporation of efficient data-based methodologies in Earth system science and climate studies necessitates a reevaluation of the integration between physical and deep learning models. This integration poses new challenges in machine learning. The objective of this project is to concentrate on well-identified key issues, allowing us to surpass current limitations and devise novel strategies for leveraging deep learning in climate studies and Earth science.

Our practical aim is to develop methods that facilitate learning the intrinsic structure of multi-variate geophysical fields. These physics-informed deep learning methods will draw insights from numerical modeling, geophysical products estimated through observations, or direct spatial observations of these fields.

References

[1] Reichstein, Markus, et al. “Deep learning and process understanding for data-driven Earth system science.” Nature 566.7743 (2019): 195-204.

[2] Wang, Jindong, et al. “Generalizing to unseen domains: A survey on domain generalization.” IEEE Transactions on Knowledge and Data Engineering (2022).

[3] Cazelles, Bernard, et al. “Disentangling local and global climate drivers in the population dynamics of mosquito-borne infections.” Science Advances 9.39 (2023).

Main Objective Summary

Leveraging Deep Learning for Generalization in Geospatial Data Analysis

A pivotal challenge hindering the effective utilization of Deep Learning (DL) methods in geospatial data analysis lies in their limited ability to generalize across diverse contexts and scenarios. Unlike physics-based approaches that capitalize on causal mechanisms embedded in a physical dynamics model, enabling adaptability to a wide range of contexts, DL models struggle with the inherent complexity of real-world dynamical systems. The discrepancy arises from the assumption of independent and identically distributed (i.i.d.) data in typical Machine Learning (ML) approaches during training and inference, a condition rarely met in climate applications. This well-known Out of Distribution problem, acknowledged in the ML community [2], has spurred numerous contributions. However, most are confined to classification and regression problems, falling short of capturing the intricacies observed in geospatial phenomena.

Furthermore, the unique demands of geospatial developments introduce additional complexities. Models trained on simulated data may not translate effectively to real-world scenarios. Similarly, a model trained on observations from one satellite within a constellation may not generalize to others, and regional or temporal variations present additional challenges.

This project aims to advance fundamental Artificial Intelligence (AI) research by addressing the problem of generalization in modeling geophysical fields. The focus will be on two prominent geoscience challenges: downscaling low-resolution sea surface height fields through integration with high-resolution sea surface temperature, utilizing observed or simulated fields, and quantitatively estimating rainfall fields from space via remote sensing satellites.

The project unfolds in three comprehensive work packages. The first centers on fundamental contributions to deep learning, specifically addressing the problem of domain generalization. Simultaneously, the other package concentrate on the practical application of these fundamental contributions, with a keen focus on downscaling sea surface height fields and estimating rainfall fields from space. Furthermore, we aim to use the climatic knowledge in understanding epidemic spreads and modeling epidemic trajectories of vector-borne diseases [3].

This strategic division between fundamental and applied contributions streamlines the project’s description, emphasizing that both aspects will be concurrently and synergistically developed throughout the course of the completion of this project.

Expected Impact: The contemporary landscape of Earth observations is characterized by increasing complexity, with the integration of satellite constellations and ensemble methods. In response to this evolution, there is a pressing need for a paradigm shift in modeling approaches to optimize the extraction of valuable insights from these intricate datasets. This proposed project significantly contributes to reinforcing crucial interdisciplinary research goal, fostering synergies between the AI and environmental and climatic sciences within the Institute.

At its core, the project addresses the formidable challenge of achieving coherent modeling of geospatial and geophysical fields. The anticipated impacts span across three critical dimensions:

Scientific Advancements: Unraveling the complexities of climate dynamics and its evolution through advanced modeling techniques will contribute significantly to the scientific understanding of Earth’s systems. This, in turn, lays the groundwork for more accurate and nuanced climate predictions.

Societal Relevance: The project holds substantial implications for societal well-being by enabling more precise modeling of unforeseen extreme events. This capability is instrumental in enhancing preparedness and response strategies, ensuring communities are better equipped to navigate and mitigate the impacts of unpredictable environmental phenomena and controlling unprecedented cases of epidemics.

Industrial Applications: Bridging the gap between AI research and environmental sciences is poised to yield practical outcomes with profound industrial relevance. The improved modeling of geophysical fields will have direct applications in weather prediction and energy forecasting, offering tangible benefits for industries reliant on accurate environmental insights.

In essence, the project’s far-reaching impacts extend beyond academic realms to directly address the evolving needs of contemporary Earth observations, fostering interdisciplinary collaborations and paving the way for advancements with transformative implications for science, society, and industry.

TEAM MEMBERS:

Dr. Uttam Kumar, Assistant Professor, IIIT Bangalore.