Bose, Avishek | Search NASA

DOE OSTI2024

Correlating Power Outage Spread with Infrastructure Interdependencies During Hurricanes

Power outages caused by extreme weather events, such as hurricanes, can significantly disrupt essential services and delay recovery efforts, underscoring the importance of enhancing our infrastructure's resilience. This study investigates the spread of power outages during hurricanes by analyzing the correlation between the network of critical infrastructure and outage propagation. We leveraged datasets from Hurricanemapping.com, the North American Energy Resilience Model Interdependency Analysis (NAERM-IA), and historical power outage data from the Oak Ridge National Laboratory (ORNL)'s EAGLE-I system. Our analysis reveals a consistent positive correlation between the extent of critical infrastructure components accessible within a certain number of steps (k-hop distance) from initial impact areas and the occurrence of power outages in broader regions. This insight suggests that understanding the interconnectedness among critical infrastructure elements is key to identifying areas indirectly affected by extreme weather events.

Bose, Avishek↗

DOE OSTI2024

Comparative Study of Large Language Model Architectures on Frontier

Large language models (LLMs) have garnered significant attention in both the AI community and beyond. Among these, the Generative Pre-trained Transformer (GPT) has emerged as the dominant architecture, spawning numerous variants. However, these variants have undergone pre-training under diverse conditions, including variations in input data, data preprocessing, and training methodologies, resulting in a lack of controlled comparative studies. Here we meticulously examine two prominent open-sourced GPT architectures, GPT-NeoX and LLaMA, leveraging the computational power of Frontier, the world’s first Exascale supercomputer. Employing the same materials science text corpus and a comprehensive end-to-end pipeline, we conduct a comparative analysis of their training and downstream performance. Our efforts culminate in achieving state-of-the-art performance on a challenging materials science benchmark. Furthermore, we investigate the computation and energy efficiency, and propose a computationally efficient method for architecture design. To our knowledge, these pre-trained models represent the largest available for materials science. Our findings provide practical guidance for building LLMs on HPC platforms.

Yin, Junqi↗