Towards Digital Twin: Introduction to Foundation Models for Geoscience

Towards Digital Twin: Introduction to Foundation Models for Geoscience

Webinar Speaker:

Sujit Roy, Ph.D., Johannes Schmude, Ph.D.



About the Webinar

Foundation models (FMs) are essentially pre-trained neural networks that serve as the underlying architecture for a range of computer vision tasks, including object classification, detection, and segmentation. Originating from the concept of building on an existing ‘foundation’, these models are constructed by training on large and diverse datasets. This training process enables them to capture a wide array of visual features applicable across multiple domains, making them incredibly versatile.

The significant advantage of foundation models lies in their ability to perform specific tasks without requiring the extensive dataset typically necessary for training a model from scratch. By leveraging the deep learning neural networks’ advanced representational learning capabilities, foundation models can generalize effectively across different tasks. This approach not only enhances the efficiency of deploying deep learning models but also makes the power of deep learning more accessible for a variety of applications. The successful development and progress of foundation models in science rely on the collective work of interdisciplinary teams, encompassing various research groups, academic and government bodies, and tech companies. By adhering to the principles of open science, such teamwork encourages transparency, the ability to replicate results, and the sharing of knowledge, thereby facilitating wider availability of datasets, models, and codes for fine-tuning. In this webinar, we will cover two FM designs, where we learn about dynamics, and large spatial and temporal scales. These models can help in understanding earth processes and advancement towards Digital Twin.


About the Speakers

Dr. Sujit Roy works as a Computer Scientist Level VI at NASA’s Interagency Implementation and Advanced Concept Teams (IMPACT), where he leads the development of foundational models for analyzing satellite imagery and enhancing weather forecasting, resulting in practical multiple scientific applications. Prior to his tenure at NASA IMPACT, Dr. Roy contributed to the field of Explainable AI at the University of Manchester. He received his PhD in Computer Science from Ulster University in collaboration with the Indian Institute of Technology Kanpur, India. In his PhD, he contributed to the domain of Computational Neuroscience by developing algorithms for Advancing MEG- and EEG-Based Decoding of Motor Imagery for Practical Brain-Computer Interfaces. He has experience of 10 years in Research and Development in the field of machine learning and deep learning. He is also a Co-founder of BrainAlive Research Pvt Ltd. His professional repertoire spans deep learning, brain-computer interfaces, and image processing. Dr. Roy is particularly focused on advancing computer vision, video/image processing, explainable artificial intelligence, signal processing/synthesis, and reinforcement learning.