Agentic AI Framework for Interactive and Scalable Sonic Data Processing
Wednesday, June 24, 2026 3:45 PM to 5:15 PM · 1 hr. 30 min. (Europe/Berlin)
Foyer D-G - 2nd Floor
Project Poster
AI Applications powered by HPC TechnologiesGeosciencesIndustrial Use Cases of HPC, ML and QCLarge Language Models and Generative AI in HPCML Systems and Frameworks
Information
Poster is on display.
Modern seismic processing workflows involve complex, multi-stage procedures spanning data quality control, conditioning, visualization, and increasingly large-scale imaging and inversion tasks. These workflows typically require deep domain expertise and manual orchestration across heterogeneous tools and execution environments, creating a significant barrier to interactivity and productivity—particularly when operating on large datasets.
This project presents an agent-driven geoscience processing platform designed to bridge natural-language user interaction with executable, high-performance data workflows. The system adopts a dual-agent architecture that was iteratively refined from an initial multi-agent sequence to reduce response latency, improve conversational interactivity, and strengthen logical coherence. The first agent operates as a Requirements Gathering Agent, engaging the user through natural language to incrementally elicit and refine processing intent while leveraging semantic retrieval over domain documentation and prior workflows. Once sufficient context is established, the expanded requirements are forwarded to a Workflow Creation Agent, which synthesizes structured execution plans represented as validated workflow graphs.
These workflows are executed through a dedicated system execution layer that integrates multiple seismic processing backends, data management services, and visualization components. The platform supports tool calling for vector store search, document retrieval, and question answering, enabling adaptive planning without hard-coded pipelines. Execution results are streamed back to the user with progressive visualization, allowing real-time inspection and iteration.
The current implementation focuses on sonic well data, providing automated quality control and conditioning for LAS and DLIS formats, including consistency checks, filtering, automatic gain control, and waveform preparation for imaging. These procedures are executed transparently without requiring users to manually configure individual processing steps, while still exposing intermediate results for inspection and validation.
Preliminary results demonstrate that consolidating requirement interpretation and workflow synthesis into two specialized agents significantly improves response time and user engagement, enabling interactive data exploration even in complex processing scenarios. The system has been validated on real sonic datasets, showing reliable execution and seamless visualization across heterogeneous tools.
Future work will extend the platform toward imaging and inversion workflows and introduce smarter planning mechanisms for terabyte-scale seismic volumes, addressing data locality, movement costs, and execution scheduling across large environments. This positions the framework as a scalable foundation for interactive, agent-driven scientific computing in geophysics.
Contributors:
Modern seismic processing workflows involve complex, multi-stage procedures spanning data quality control, conditioning, visualization, and increasingly large-scale imaging and inversion tasks. These workflows typically require deep domain expertise and manual orchestration across heterogeneous tools and execution environments, creating a significant barrier to interactivity and productivity—particularly when operating on large datasets.
This project presents an agent-driven geoscience processing platform designed to bridge natural-language user interaction with executable, high-performance data workflows. The system adopts a dual-agent architecture that was iteratively refined from an initial multi-agent sequence to reduce response latency, improve conversational interactivity, and strengthen logical coherence. The first agent operates as a Requirements Gathering Agent, engaging the user through natural language to incrementally elicit and refine processing intent while leveraging semantic retrieval over domain documentation and prior workflows. Once sufficient context is established, the expanded requirements are forwarded to a Workflow Creation Agent, which synthesizes structured execution plans represented as validated workflow graphs.
These workflows are executed through a dedicated system execution layer that integrates multiple seismic processing backends, data management services, and visualization components. The platform supports tool calling for vector store search, document retrieval, and question answering, enabling adaptive planning without hard-coded pipelines. Execution results are streamed back to the user with progressive visualization, allowing real-time inspection and iteration.
The current implementation focuses on sonic well data, providing automated quality control and conditioning for LAS and DLIS formats, including consistency checks, filtering, automatic gain control, and waveform preparation for imaging. These procedures are executed transparently without requiring users to manually configure individual processing steps, while still exposing intermediate results for inspection and validation.
Preliminary results demonstrate that consolidating requirement interpretation and workflow synthesis into two specialized agents significantly improves response time and user engagement, enabling interactive data exploration even in complex processing scenarios. The system has been validated on real sonic datasets, showing reliable execution and seamless visualization across heterogeneous tools.
Future work will extend the platform toward imaging and inversion workflows and introduce smarter planning mechanisms for terabyte-scale seismic volumes, addressing data locality, movement costs, and execution scheduling across large environments. This positions the framework as a scalable foundation for interactive, agent-driven scientific computing in geophysics.
Contributors:
Format
on-demandon-site
