🤖 machine learning

OpenSeisML: Open Large-Scale Real Seismic and well-log Dataset for Generative AI

This paper introduces OpenSeisML, a curated collection of real seismic and well-log datasets from the UK National Data Repository, featuring an automated pipeline for time-to-depth conversion, to address the scarcity of public data and enable the training of generative AI models for seismic inversion and uncertainty quantification.

Original authors: Ipsita Bhar, Huseyin Tuna Erdinc, Thales Souza, Charles Jones, Felix J. Herrmann

Published 2026-05-21

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Ipsita Bhar, Huseyin Tuna Erdinc, Thales Souza, Charles Jones, Felix J. Herrmann

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to teach a robot chef how to bake the perfect cake. To do this, the robot needs to taste thousands of real cakes to understand how flour, sugar, and eggs interact. However, in the world of oil and gas exploration, the "cakes" are underground rock formations, and the "recipes" are hidden inside private company vaults. Most of the high-quality data needed to train these AI chefs is locked away, leaving researchers to practice on fake, computer-generated cakes that don't quite taste like the real thing.

This paper introduces OpenSeisML, a new, open-source "kitchen" filled with real, high-quality seismic data and well logs, designed specifically to train the next generation of AI for finding oil and gas.

Here is a breakdown of what they did, using simple analogies:

The Problem: The "Fake Cake" Dilemma

For a long time, researchers have tried to train AI to figure out what's underground using synthetic data (computer simulations).

The Compass Model & SEAM: These are like very detailed, realistic-looking fake cakes. They look good, but there's only one version of each. If you train your AI on just one fake cake, it memorizes that specific cake and fails when it sees a slightly different real one.
OpenFWI: This is like a massive library of fake cakes with different shapes. While there are many of them, they are still made in a controlled lab. They lack the messy, chaotic "flavor" of real geology.

The result? AI models trained on these fakes often struggle when they encounter the messy reality of the actual Earth.

The Solution: The "Real Cake" Library

The authors went to the UK National Data Repository (NDR), a government-run public library of real-world geological data. They built an automated pipeline to grab real seismic surveys (3D images of the underground) and real well logs (measurements taken from actual drill holes).

Think of this pipeline as a super-efficient assembly line that cleans and prepares these raw ingredients so an AI can eat them immediately.

The Assembly Line: How They Prepared the Data

The paper describes a four-step process to turn raw data into training material:

Gathering the Ingredients (Data Collection):
They downloaded massive 3D seismic datasets and associated well logs from the UK. They were picky, mostly choosing data from after the 1990s because older data was like "expired ingredients"—messy, inconsistent, and hard to use automatically.
Aligning the Map (Coordinate System):
Imagine trying to overlay a satellite photo of a city with a hand-drawn map of a subway system. If the scales don't match, the lines won't line up. The researchers made sure all the seismic data and the well logs were on the exact same coordinate map so the AI knows exactly where the drill hole is relative to the seismic image.
The Time-Travel Conversion (Time-to-Depth):
This is the trickiest part. Seismic data often comes in "time" (how long it took a sound wave to bounce back), but drill holes measure "depth" (how many meters down).
- The Analogy: It's like trying to convert a recipe that says "cook for 10 minutes" into "cook until the center reaches 200 degrees." You need a conversion chart.
- The Method: They used "checkshot" data (measurements that tell you exactly how long sound takes to travel to specific depths) to build a smooth velocity model. Think of this as creating a "speed map" of the underground. They used a mathematical tool called Radial Basis Functions (RBF) to fill in the gaps between the drill holes, creating a smooth, 3D speed map that acts as a translator between time and depth.
Slicing the Loaf (Creating Training Data):
Once the 3D data was converted to depth, they sliced it into 2D strips that passed right through the drill holes. They then resized these slices to a standard size (like resizing all photos to 1080p) so the AI could learn from them consistently.

The Result: A New AI Chef

The researchers tested this new dataset by training a Diffusion Model (a type of Generative AI that learns patterns to create new images).

The Test: They trained the AI on just 40 real wells.
The Outcome: The AI successfully learned the statistical "flavor" of the underground. It could generate new, realistic velocity models that looked very similar to the real ground truth.
The Goal: The ultimate aim is to use this AI to create thousands of "what-if" scenarios. This helps geologists understand the uncertainty of what lies underground, rather than just giving them one single guess.

Summary

In short, the paper says: "Stop training your AI on fake, single-version simulations. We have built an automated pipeline to harvest, clean, and organize real seismic data from the UK. This allows AI to learn the true, messy, statistical nature of the Earth, leading to better predictions for seismic inversion."

They are currently working with 2D slices and plan to scale this up to 3D and include more wells (up to 1,000) to make the AI even smarter.

Technical Summary: OpenSeisML

Problem Statement
The development and evaluation of Generative AI (Gen-AI) workflows for seismic inversion are currently hindered by a scarcity of realistic, high-quality velocity models. While machine learning has accelerated seismic inversion by reducing the computational cost of iterative methods, most available high-quality data are privately owned by oil and gas companies. Existing public benchmarks, such as the Compass model, SEAM, and OpenFWI, suffer from significant limitations: they are often restricted to single deterministic realizations, lack stochastic variability, or rely on synthetic data that fails to capture the full geological complexity and variability of real field conditions. Consequently, models trained on these datasets risk overfitting to specific geologies or synthetic setups, struggling to generalize to diverse subsurface environments.

Methodology
To address these gaps, the authors present OpenSeisML, an automated data curation pipeline designed to generate large-scale, real seismic and well-log datasets from the UK National Data Repository (NDR). The workflow consists of four primary components:

Data Collection: The pipeline utilizes the UK-NDR portal to bulk download 3-D migrated seismic data (SEG-Y) and associated well data (LAS and checkshot files) via APIs. To ensure quality and reproducibility, the authors prioritize surveys acquired after the 1990s to avoid inconsistent headers and uncertain well coordinates, though a small number of pre-1990 surveys were included after verification.
Preprocessing and Alignment: Seismic volumes and well data are aligned within a common coordinate reference system (CRS). To represent 3-D seismic data on a regular grid, the authors estimate survey boundaries using a concave hull to account for irregular acquisition geometries and extract the largest contiguous rectangular region with regular grids.
Time-to-Depth Conversion: Since post-stack seismic data may be in the time domain while well logs are in depth, a conversion is necessary. The authors utilize checkshot data (providing depth and two-way travel times) to compute average velocities. These velocities are spatially interpolated using a Radial Basis Function (RBF) to construct a smooth 3-D average velocity volume. This volume is used to convert time-migrated seismic data to the depth domain. The authors explicitly note that the goal is a consistent time-depth mapping rather than a highly accurate interval velocity model, aiming to capture the overall statistical behavior of the velocity field.
Training Data Generation: The final step involves extracting 2-D depth-domain seismic sections along "quasi-2D lines" that pass through selected well trajectories. These sections are resampled to a $256 \times 512$ grid using a 2D FFT with a smooth low-pass filter to harmonize resolutions between the seismic traces and the typically finer-sampled well logs. The final dataset comprises 2-D seismic sections integrated with well control, stored in HDF5 format.

Key Contributions

OpenSeisML Dataset: A curated collection of real seismic datasets from the North Sea basin, designed specifically to support Gen-AI workflows for seismic inversion.
Automated Pipeline: A reproducible, automated workflow for data curation that handles the complexities of aligning real-world seismic and well data, including time-to-depth conversion using checkshot-derived velocity models.
Statistical Representation: Unlike previous benchmarks that offer single realizations, this dataset captures the statistical distribution of subsurface properties, enabling the synthesis of multiple statistically consistent realizations for uncertainty quantification.
Multi-Parameter Support: The dataset includes not only velocity but also petrophysical measurements (gamma-ray, neutron, density, sonic, and resistivity logs), providing constraints on lithology, porosity, and fluid properties.

Results
The authors conducted preliminary experiments training a diffusion-based generative model on the curated OpenSeisML data.

The model was initially trained on the Compass dataset for baseline comparison and then retrained using data from only 40 wells from the OpenSeisML collection.
The results demonstrated that the model could generate realistic velocity model distributions that closely align with ground truth data derived from the wells.
The authors note that these results are preliminary; future work involves scaling the training to at least 1,000 real well logs to further improve accuracy and generalization.

Significance and Claims
The paper claims that OpenSeisML addresses the critical bottleneck of data scarcity in the development of trustworthy and reproducible ML models for geophysics. By leveraging real field data rather than synthetic or deterministic models, the framework aims to capture the true statistical characteristics and geological complexity of subsurface conditions. The authors position this work as a foundational step toward training generative models that can act as priors for seismic inversion and facilitate uncertainty quantification. They emphasize that the current evaluation is in a 2-D setting, with plans to extend the framework to 3-D, and highlight the potential for these datasets to evaluate multi-parameter inversion algorithms in the future.