AASPI - University of Oklahoma

Seismic ML is powerful.
But did it actually learn GEOLOGY?

A multi-attribute seismic ML uncertainty assessment for geoscientists

Multi-attribute seismic machine learning gives us remarkable new ways to interrogate seismic data. The outputs can be stunning. Facies boundaries appear in minutes. Waveform patterns nobody noticed suddenly organize into what looks like coherent geology.

But here is what we've learned at AASPI these methods find patterns in whatever you give them. Feed them acquisition footprint and they will cluster your footprint beautifully. Give them unnormalized attributes and the result reflects data scaling, not geology. Skip the despiking step and one bad trace pulls a cluster center to the wrong place. Run a SOM on a full volume when you care about a 50ms window and the method spends most of its effort distinguishing water column from basement.

None of these are software problems. They are workflow problems. And they are extremely common - in student projects, in industry practice, and in published research - regardless of what software produced the result. A good interpretation is one you can explain and defend. The tool that generated it is secondary.

"The algorithm found exactly what you gave it. The question is whether what you gave it was geology."

- the whole point of this questionnaire!!!

This tool was developed at AASPI (Attribute Assisted Seismic Processing and Interpretation) at the University of Oklahoma — a research group whose philosophy has always been to understand what the data is telling us, not just to produce a result. This questionnaire is that philosophy in checklist form. It is designed to work with any multi-attribute seismic ML software or workflow. The questions apply whether you wrote the code yourself, used a commercial platform, or anything in between.

At the end you get two uncertainty matrices - one for your data and attribute quality, one for your method rigor and geologic validation - plus specific, actionable recommendations for every gap the questionnaire finds.

Methods covered:

k-means
PCA
ICA
GMM
SOM
GTM
Dimension reduction
Other unsupervised
📡
Seismic data quality
S/N, footprint, migration, bandwidth, phase, spectral balancing
🎛️
Attribute preparation
Selection, redundancy, normalization, cyclical transforms, despiking
📐
Dimension reduction
PCA vs ICA, component evaluation, footprint in components
🧮
Unsupervised classification
Method choice, windowing, parameterization, stability, calibration
🔍
Explainability
Attribute influence, spatial consistency, sensitivity testing
🗺️
Geologic validation
Artifact checks, analog comparison, well calibration, uncertainty communication
HB
Dr. Heather Bedle - Dr. David Lubo-Robles - Dr. April Moreno-Ward - Hilmi Putra , University of Oklahoma
AASPI (Attribute Assisted Seismic Processing and Interpretation) - School of Geosciences

AASPI is a research group, not a software vendor. Our tools are built to be understood, not just run. This questionnaire grew out of years of watching interpreters trust ML outputs without asking hard questions — and wanting to give them a structured way to do better, whatever software they use.

If you hit a failure mode this tool doesn't cover, find a question that doesn't fit your workflow, or have ideas for what to add - I genuinely want to hear from you. Feedback, suggestions, and war stories all welcome.

aaspi@ou.edu  ·  aaspi.ou.edu
~15-20 min - 7 sections - printable summary

Version 1.0 - 2025. Covers unsupervised volumetric classification and dimension reduction. Works with any multi-attribute seismic ML software or workflow. Supervised methods coming in a future version. Your answers are never stored or transmitted.