I Asked GPT to Generate Seismic Lines. This Is What Happened

AI generated salt dome. It looks convincing at a glance.

Artificial intelligence can write essays, plan trips, and debate philosophy. But I had a more specific, geophysical question: Can it create a believable seismic section?

I didn't want a cartoon or a diagram. I wanted an image that would make an interpreter do a double-take. So, I gave a generative AI a simple geological sketch, a half-graben with a listric fault, and asked it to generate a synthetic seismic line.

The results were fascinating, deeply flawed, and surprisingly instructive.

The AI's Seismic Gallery: A Geologist's Uncanny Valley

At first glance, the outputs are undeniably impressive. The AI has clearly been trained on thousands of seismic images and has learned the visual "grammar."

Another AI-generated salt dome.

AI generated "rift graben with salt and pull-up effect". It mimics the jargon.

"Compressional thrust belt?" The AI struggles with complex structural styles.

The Geoscientist's Autopsy: Where the AI Fails

While visually stylized, these images collapse under geophysical scrutiny. The AI is an artist, not a physicist. Its mistakes are systematic:

• The "Reflectors" are Lies: They don't represent wave propagation or travel times. They are simply lines drawn to mimic the look of continuity.
• No Rock Physics: The contrast between black and white isn't based on acoustic impedance. It's an aesthetic choice.
• Geological Nonsense: Structures are often kinematically impossible. You'll see extensional and compressional features merged in ways that violate basic mechanics.
• Ghosts of Diffractions: Diffraction patterns appear randomly, unrooted in actual point sources like fault tips.
• The Scale is Meaningless: There is no vertical scale (time or depth) because there is no velocity model to create one.

In short, the AI creates a convincing facade. It's a seismic caricature, it captures the essence of the style but none of the substance.

The Real Danger: The Illusion of Competence

This experiment revealed a danger far greater than AI being wrong: the danger of it being just right enough to be persuasive.

To a non-expert, or even an expert in a hurry, these images scream "seismic data." They use the right visual vocabulary. This is the true risk. An AI could generate a "plausible" seismic line to support a flawed prospect for a rushed presentation, bypassing the rigorous physics that keeps our interpretations honest.

The Core Truth: Interpretation is Physics, Not Art

This brings me to the most important point, one this experiment burned into my mind:

Interpretation is not drawing.
Interpretation is not preference.
Interpretation is not authority.

Interpretation is a testable hypothesis about subsurface geometry. A valid interpretation must be able to, through the laws of physics, generate the seismic response we actually recorded. If your proposed structure cannot recreate the data, it is not a different "opinion", it is scientifically false.

Too many good ideas are rejected because they "looked weird" to a manager, not because they violated physics. Familiarity, not fidelity, becomes the judge. This is the enemy of discovery.

What's Next: Replacing Illusion with Reality

This AI exercise wasn't just a critique. It was an inspiration. If an AI without physics can produce such intriguing fakes, what could we do with the physics?

This has convinced me to start a series on forward seismic modeling. We will:

• Take geological sketches.
• Build realistic velocity models.
• Use Python to generate synthetic seismic with actual wave-equation algorithms.
• Compare interpretations using real, quantitative misfit measures.

Two interpreters may draw different faults, but the Earth only has one. The wavefield is the ultimate judge, and we can now use code to act as its jury.

A Hopeful Conclusion

Asking an AI to generate seismic was not a game. It was a revelation. It forced a clearer, more rigorous way of thinking about the subsurface.

And I'll admit, I smiled when I saw it. I smiled because even an AI that knows nothing about geology instinctively tries to imitate physics. It knows that's the soul of the image.

Our job as geoscientists isn't to draw what we think is there. It's to propose what must be there, and then let the unyielding rules of physics have the final say. The future isn't in AI that draws seismic, but in AI that helps us obey its physical laws.

Welcome to The Geoscientist Blog. More experiments soon.