Skip to content
Daniel Wamriew
Projects
Research

Conditional Diffusion Models for Microseismic Localization

Generative AI for a classic geophysical inverse problem: locating microseismic events directly from noisy waveform gathers, without arrival-time picking, and with uncertainty estimates from repeated sampling.

Illustrative
Problem
Locating microseismic events from noisy recordings is an ill-posed inverse problem. Travel-time methods depend on accurate arrival picks and velocity models, and deterministic deep learning returns a single answer with no measure of reliability.
Approach
A conditional diffusion model that generates the event location as two Gaussian heatmaps (X–Z and Y–Z), conditioned on features extracted from the recorded three-component waveforms.
My contribution
Formulated event localization as conditional heatmap generation, combining waveform preprocessing, learned features and quantitative baseline evaluation, and investigated noise sensitivity and repeated stochastic sampling to characterize uncertainty.

Overview

Microseismic monitoring tracks the small seismic events released as rock fractures, for example during hydraulic fracturing or CO₂ injection. Each event has to be located from the waveforms recorded by a sensor array. This project, with Tariq Alkhalifah at KAUST, replaces that inversion with a conditional diffusion model (CDM): given a waveform gather, it generates the event location directly, without manual arrival-time picking, and its repeated samples give a measure of how certain the location is.

The work was developed on the geometry and recordings of the Tony Creek Dual Microseismic Experiment (ToC2ME) in Alberta, Canada, a shallow-borehole array that monitored hydraulic fracturing.

Problem addressed

Classical localization methods each carry a weakness. Travel-time inversion depends on the quality of the arrival picks and the velocity model; in low signal-to-noise conditions, small picking errors can move a location by tens of meters. Migration and full-waveform methods are more robust to picking but computationally heavy. Deep learning regressors are fast but return one deterministic answer, with no indication of when that answer should not be trusted.

Motivation

A generative model does not have to commit to a single estimate. Trained to produce locations conditioned on the data, a diffusion model can be sampled many times, and the spread of its samples becomes an uncertainty estimate. The question was whether that benefit could be obtained without giving up accuracy or speed.

Technical approach

  • Data. Synthetic events were placed in the volume where most ToC2ME events occurred, modeled on the site's velocity model, and combined with noise recorded in the field, giving realistic semi-synthetic waveform gathers for training and testing.
  • Removing the origin time. A cross-correlation step removes the unknown event origin time, so the network sees the moveout pattern of the wavefront rather than absolute times.
  • Locations as heatmaps. Diffusion models work best on distributions rather than points, so each 3-D location is represented as two Gaussian heatmaps, one in the X–Z plane and one in the Y–Z plane.
  • Conditional diffusion. The model learns to turn noise into a location heatmap, guided by features extracted from the recorded waveforms. The location is read from the heatmap peaks.
  • Uncertainty. Sampling the same gather many times gives a spread of locations, which serves as an uncertainty estimate.

My contributions

  • Formulated localization as conditional heatmap generation and designed the dual-projection representation.
  • Built the semi-synthetic data pipeline and the preprocessing.
  • Implemented, trained and evaluated the model, including the CNN baseline, noise sensitivity, sampling-step sensitivity, uncertainty quantification and the field example.
  • Wrote the manuscript.

Results

On the 1,774 test samples:

Metric X Y Z
MAE (m), CDM 18.0 12.4 21.5
MAE (m), CNN baseline 28.5 32.4 28.7
R², CDM 0.929 0.984 0.816
R², CNN baseline 0.988 0.990 0.936
  • Against a CNN baseline trained on the same data, the CDM has lower MAE on every axis. The CNN has higher R² and lower error standard deviation, reflecting a narrower spread of predictions.
  • Noise. Localization errors stay below 30 m up to about 70% noise relative to the maximum signal amplitude. In the comparison across noise levels, the CNN's error reaches nearly 200 m at 50% noise, while the CDM's grows more gradually.
  • Uncertainty. Over the test set, the spread of the repeated samples correlates with the actual mislocation (Pearson r = 0.98 overall, 0.35 after removing the 5% largest errors), so the ensemble flags the predictions most likely to be wrong.
  • Field data. Applied without retraining to a real ToC2ME event (Mw 2.05), the model placed it 11.5 m from the catalog location. This is a single-event proof of feasibility, not a statistical validation.

Lessons learned

  • Representation matters. Turning points into heatmaps is what lets a diffusion model, built for distributions, solve a coordinate problem.
  • Report the trade-offs. The CNN wins on R² and spread; the CDM wins on mean error and robustness. Both belong in the comparison.
  • Limits are explicit. The model is trained on a simplified velocity model and does not enforce wave physics. Richer training data and fine-tuning on labeled field data are the next steps.