In 2024, ThinkOnward launched the first version of its Geophysical Foundation Model (GFM) and has made it open source. Why? Because the power of community will advance the GFM faster than any company can and take it to new places even we couldn’t predict.

Our mission is to accelerate AI innovation in geoscience through:

• Community-driven development
• Cutting-edge machine learning techniques
• Accessible, powerful AI tools for subsurface exploration

The ThinkOnward Geophysical Foundation Model (GFM) is based on important recent work in foundation model development (Sheng et al., 2023) and addresses some issues in that work. Good foundation models are crucial in developing more accurate or “smarter” artificial intelligence (AI).

Released in 2024, technically, the GFM is a Vision Transformer Masked Autoencoder (ViT-MAE) (He et al., 2021) model that has been pre-trained on seismic data using a novel trace-based masking approach. For more information about the ThinkOnward GFM, see our previous blog post “Breakthrough: New foundation model accelerates machine learning for geophysical applications.”

Why does open source matter?

ThinkOnward made the bold decision to make our GFM open source. This model is another valuable resource that’s part of our community platform, which empowers an active network of engineers and scientists (geoscientists, data scientists, and many other disciplines) to solve problems in the subsurface, natural resources, and energy domains.

Since its start, ThinkOnward has crowd-sourced solutions to industry problems by sponsoring challenges, bounties, and projects, which people all over the world are invited to solve. Challenges are based on real-world problems, offer cash prizes for the best solutions, and are further explained below in this post.

Crowd-sourcing results in a variety of approaches and great new ideas, as well as bringing the thinkers of those ideas into the community. In some cases, Challenge prize winners have become consultants on client projects.

With this philosophy as the foundation of our company, making the GFM available as open-source code was the only way to go.

Free resources to implement and use the GFM

However, the GFM isn’t just “out there.” We have many resources to help people get started and use it to solve challenges—such as official ThinkOnward Challenges or their own real-world problems.

Webinars and Tutorials

Recent webinars gave an overview of the GFM and provided a step-by-step tutorial on how to download and set up the GFM in your IT environment. For recordings of these webinars, click these links:

• Breakthrough: New foundation model accelerates machine learning for geophysical applications (Blog post)
• Geophysical foundation model; open source and open weights (November 2024)
• Geophysical foundation model usage & installation (January 2025)

Global community GFM-related challenges

ThinkOnward Challenges present real-world problems to the global community to solve for prize money, recognition, and, in some cases, possible future opportunities to work on customer projects. Challenges are a great way to initially engage with and understand the technology and how it can be applied.

For each challenge, we define the problem, working parameters, data sets, guidelines, prize amounts, and success criteria. The links in the list below of recent GFM-related challenges provide all specific details—including the winners of the first two listed.

• In the Patch the Planet Challenge, participants were tasked with developing innovative techniques to interpolate missing seismic traces from 3D seismic data. We developed the GFM during this challenge by pre-training a ViT-MAE model on 450 synthetic seismic data volumes generated using Synthoseis (Merrifield et al., 2022).
• For the Image Impeccable Challenge, participants were provided a series of “very noisy” 3D seismic volumes. They were required to build a model that could intake a seismic volume and denoise the volume in an accurate and efficient manner. The GFM was used as a baseline model for the prediction leaderboard to give participants a model performance target.
• The Dark Side of the Volume Challenge tasked participants with identifying faults on seismic volumes and building a 3D polygon around the faults; all solutions were required to use machine learning. The data for this challenge were 400 3D seismic volumes and corresponding fault labels, and a test dataset that participants used to make predictions, which were submitted for scoring. (This challenge is closed, but results have not been finalized.)

Who’s participating?

Challenge participants include seasoned industry practitioners and students in computer science, data science, and geoscience from around the world. 

A recent winning submission came from a team of medical imaging professionals whose primary job is analyzing MRIs and CT scans of patients’ bodies. They adapted their medical diagnostic techniques to the seismic data posted for a challenge.

Their solution used creative new techniques that worked effectively across many different terrains and geological conditions, maintaining a high performance regardless of the geological environment. This kind of broad, international, cross-industry participation is already adding significant insight and value to the GFM.

Get access to the latest GFM Updates

As part of our open-source community, you can access any updates we publish to the GFM. In a recent webinar, we mentioned we had just released some new segmentation code, and a webinar attendee commented in the chat that they had taken the code and were already using it for classification segmentation. That’s the kind of action we love to see!

Join the GFM party!

• Register for our next webinar coming on Wednesday, March 12:
  • Enhancing Seismic Clarity: Open Source AI Denoising: We'll demonstrate how to remove seismic multiples using a GFM, similar to the Image Impeccable Challenge, fine-tuned on synthetic seismic data. Explore the potential of open-source weights and code, coming soon to HuggingFace and GitHub.
  • Register today

• To join our community, do one or more of these:
  • Register online and sign up for a Challenge
  • Download the GFM from Huggingface (an AI community at https://huggingface.co/) using the detailed instructions provided or in this recorded webinar

References

• He, K. et al. (2021). Masked Autoencoders Are Scalable Vision Learners.
• Merrifield, T.P., et al. (2022). Synthetic Seismic Data for Training Deep Learning Networks, Interpretation, 10(3), SE31-SE39.
• Sheng, H. et al. (2023).Seismic Foundation Model (SFM): a new generation deep learning model in geophysics.