by the GANSTERS
model2seismic is a system for rapid forward modelling and seismic inversion in one. seismic2model is an equally valid name. It uses Generative Adversarial Networks GANs, to map between velocity and the seismic domain in both directions. Once trained on appropriate data, the system understands how to translate between the two domains, can generalise to datasets beyond the training space. It Rocks!
The results shown on this page were generated by a system built and trained at the Agile Subsurface Hackathon 2017 in Paris. Since the event, the team has trained the system further leading to more stunning results and two publications:
Training is expensive and it takes skill, careful preparation of training data and patience to get these complex systems to converge and generalise. But once trained a systemcna run n inference pass in seconds. That means seismic inversion or forward modelling applied in seconds.
So far we have trained sythetic data made with simple forward modelling. Yet the system performs well on seismic data with unknown sources. The training set can be expanded to include more geology and realistic physics. Multiple models can be trained to target particular processing workflows.
We are not actively developing a software solution for this. The best way to access the tech is to access the know how, most of the team are availale for consulancy or hire. Contact us.
The model was trained on 200x200 greyscale velocity models and synthetic seismic models created by 1D convolution with a Ricker wavelet of frequency 10Hz. These are examples from the training dataset.
The “industry standard” Marmousi model was used as a test case, to validate our models ability to generalise for geometries that it has not seen before. (Image size 2000 px) Notice that very large and arbitrary shaped domains can be mapped. Processing time 5 seconds.
The “industry standard” Marmousi model was used as a test case, to validate our models ability to generalise for geometries that it has not seen before. (Image size 2000 px) Notice that very large and arbitrary shaped domains can be mapped. Processing time 5 seconds.
This is real seismic data plucked for the web, we do not know how this was imaged nor does the system!
The follow results are generated from a system re-trained using the same "industry standard" on the Marmousi model dataset. (Huge shout to Wouter Kimman who did the retraining work)
GANsters was the name of our team at Agile Subsurface Hackathon in Paris where we won the prize for execution. We are not a company and we are not actively developing a product based on the system we have developed.
We are a group of independent researchers & consultants. We are actively working in Data Science and Machine Learning in geoscience. Impressed and interested by what you see here? then access the know how of people who put this together.
We're available for hire and consulting (subject to availability). We're independent but happy to team up. Reach out to any of us.