Multi-Strategy Approaches for Efficient Seismic Inversion on Next-Generation Supercomputing Architectures

• Grant number: 26/10717-0
• Support Opportunities: Scholarships abroad - Research Internship - Master's degree
• Start date: August 21, 2026
• End date: December 20, 2026
• Field of knowledge: Physical Sciences and Mathematics - Computer Science - Computer Systems
• Principal Investigator: Hermes Senger
• Grantee: Raphael Alexsander Prado dos Santos
• Supervisor: Martin Schreiber
• Host Institution: Centro de Ciências Exatas e de Tecnologia (CCET). Universidade Federal de São Carlos (UFSCAR). São Carlos , SP, Brazil
• Institution abroad: Université Grenoble Alpes (UGA), France
• Associated to the scholarship: 25/21092-8 - Multi-Strategy Approaches for Efficient Seismic Inversion on Next-Generation Supercomputing Architectures, BP.MS

Abstract

Full Waveform Inversion (FWI) is a computationally demanding inverse problem in which the physical parameters of a medium are estimated from accurate measurements of acoustic wave propagation. Widely adopted in the Oil and Gas industry for subsurface exploration, FWI requires solving partial differential equations (PDEs) millions of times. When implemented with finite differences stencils, these PDE solvers follow stencil computation patterns that incur intensive memory traffic, posing significant challenges for performance optimization, particularly on GPUs. Industrial-scale FWI demands continue to grow, driven by the need for higher-resolution imaging and the incorporation of more realistic physical models, which substantially increase both memory requirements and floating-point operations. As a result, state-of-the-art FWI workflows often run on clusters delivering tens of petaFLOPs for weeks or months. Several high-performance computing techniques have been proposed to mitigate these challenges, including checkpointing (reducing memory requirements for gradient computation), domain decomposition (enabling distributed execution across accelerators), and data compression (reducing storage and communication costs). However, there is a literature gap regarding integrating these three techniques at once, including trade-offs when integrating them on heterogeneous systems composed of CPUs and GPUs. This work addresses this gap by systematically exploring the integration of checkpointing, compression, and GPU-based domain decomposition for FWI on heterogeneous supercomputing clusters. Each strategy has its own configurable parameters that affect performance and involve tradeoffs that must be balanced. Additionally, combining these strategies creates interference among them. This approach requires carefully coordinated integration of these strategies so that it exploits next-generation supercomputing architectures to deliver more efficient and scalable FWI workflows for real-world applications. (AU)