Integrating advanced artificial intelligence (AI) into geoscience represents a pivotal moment, redefining how we approach exploration and interpretation of the earth's subsurface. Generative AI methods, such as large language models (LLMs), diffusion models, and physics-informed learning, offer new ways to simulate, invert, and interpret seismic data. LLMs are increasingly used in various seismic tasks ranging from interpolation and denoising to direct inversion for subsurface properties. Promising attempts have been made to develop foundational models that treat poststack seismic data like natural images. Prestack causality-aware and spatially aware foundational models have not yet been explored extensively. Diffusion models that draw samples from a learned distribution enhance data sets by generating synthetic subsurface models, filling data gaps, and creating plausible scenarios that support testing and validation. Physics-informed AI bridges the gap between empirical machine-learning approaches and traditional physics-based methodologies. Agentic AI — an emergent field leveraging the autonomous capabilities of LLMs for geophysical tasks — further expands the geophysicist's toolkit for seismic processing and workflow automation.

Introduction, AI, artificial intelligence, generative AI, language models, physics informed AI

The diverse range of contributions in this special section of The Leading Edge underscores the depth and breadth of AI's impact on the field. The section opens with the emerging theme of multimodality for vision-language interaction tailored to the context of geophysics. It closes with the promising concept of an agentic AI system for seismic processing, inspiring broader energy industry applications. Among the contributions, readers will also find diverse geophysical applications of transformer and other architectures in image-to-image and text-to-text modes, paving the way to the geophysical foundation models as well as seismic simulations and inversion using physics-informed AI.

In the special section's first paper, Aseev et al. investigate how vision-language CLIP models, optimized with contrastive learning, provide novel insights for seismic image analysis. Seismic data analysis can now directly leverage natural language textual geologic descriptions — a new direction in AI-assisted interpretation.

Pham et al. propose self-supervised pretraining of a representation model over massive seismic image volumes and demonstrate its value in accelerating multiple seismic interpretation tasks. The diversity of the training data set and the authors' new patch-as-word input approach lead to stunning quality outputs, taking seismic interpretation to a new level.

Hou et al. explore the transformative potential of generative AI for geologic document processing and discuss the challenges associated with knowledge retrieval from a diverse corpus of geoscientific documents. The work demonstrates how text and vision models enhance document processing by improving text analysis, table extraction, and figure classification.

Masaya analyzes the capabilities of popular text-to-text language models in simulating saturated rock properties. The study explores AI-based methodology choices and the error statistics when tackling the task with GPT-4 turbo, GPT-4o, GPT-4 mini, or Claude 3.5 Sonnet.

Nakata et al. develop an unconventional method for fast simulation of signals from natural earthquake events. The model simulates the physics of wave propagation and geology within the localized geologic region utilizing a subset of field data rather than numerical simulations.

Liu et al. develop a new approach to predict time-lapse velocity models for seismic monitoring. The authors' convolutional neural network with simultaneous quantile regression estimates velocity uncertainty maps and models directly from the input seismic data.

Wrapping up the special section, Kanfar et al. combine LLMs and conventional seismic operations by designing an agentic AI system to manipulate tools from an open-source Madagascar package. The result is the industry's first guardrails-equipped AI agent that performs seismic processing operations with zero code input from the user.

These diverse contributions illustrate how generative and physics-informed AI provide researchers with tools to automate complex processes, improve the fidelity of synthetic data sets, and tackle challenges in environments where conventional methods may struggle. Together, these advances are opening new pathways for exploration, enabling geoscientists to address questions with greater precision and creativity by addressing multiple geophysical problems, from seismic interpretation to data-driven simulation and uncertainty quantification.

We sincerely thank the authors for their contributions, which collectively highlight the expanding role of advanced AI in geophysical research. We are also grateful to the reviewers for their detailed and thoughtful feedback, which has profoundly impacted the quality of the publications. The geophysical community is rapidly advancing the field of AI-driven applications, and we believe this special section highlights its cutting edge.