DeepTrace Usage Guide

This guide will help you understand what kinds of DeepTrace models exist, how to use them to produce picks, and how to train new models. This guide should be read after the Installation Guide.

Back to Docs

Table of Contents

Understanding DeepTrace Models

DeepTrace’s primary function is to produce first break picks from seismic data. There are broadly four categories of DeepTrace models: smooth, sliding, moveout, and non-moveout:

Models Moveout Non-Moveout
Smooth Moveout Smooth Smooth
Sliding Moveout Sliding Sliding

Moveout vs Non-Moveout

DeepTrace expects seismic data to be presented to it in one of two ways: without any processing, or with a kind of flattening that we call a “moveout” applied.

When no processing is done to the seismic traces, and they are shown to DeepTrace as they are recorded in the field, shot-by-shot, line-by-line, a Non-Moveout model should be used. Here is an example of some non-moved-out seismic with DeepTrace picks:

Non-Moveout Smooth

If we apply a moveout trend, which applies a time-shift as a function of trace offset to flatten the data, it looks like so:

Moveout Smooth

Due to various technical reasons, DeepTrace can often perform better on data that has been moved-out. We recommend defining a moveout trend as the first step in picking a new survey. If you suspect that your data has strong anisotropy, use the “azimuthal moveout trend” tool in Phoenix. More information about defining a moveout trend in Phoenix can be found in the Phoenix Startup Guide.

Smooth vs Sliding

DeepTrace models can produce picks on many traces at once, or one at a time. These two methods are called smooth and sliding, respectively.

Smooth models predict arrival times on multiple traces simultaneously. Because all the traces are picked together in this method, the models have learned to tightly correlate the times, so the arrival wavefront ends up looking very “smooth”, hence the name. Because smooth models predict arrival times on many traces at the same time, these models are quite fast, in fact about 12x faster than sliding models. There are edge effects with sliding models, since traces at the edge of each “image” do not have as much context available to them as the traces in the center. The picks in the moveout explanation images above are from smooth models.

Sliding models create an “image” of seismic data for each and every trace separately. Because of this, they are much slower than smooth models, but tend to track fine details from trace to trace much more readily. We recommend using a sliding model to produce a final set of picks which track wavelet events with a high level of precision. The following shows picks created using a sliding model:


When deciding between using a smooth and sliding model, first consider how accurately you want the picks to track the events, how much compute/ time you have available, and at what stage of the picking process you are in. We recommend first using a smooth model to get a rough set of picks, using those to create an intiial tomo model, then using that tomo model to create a much more accurate moveout. Using the tomo moveout, finally create picks with a sliding model. This process is detailed in the Phoenix Startup Guide.


Beta Feature

Neural network performance increases with data size and variety. As an experimental feature, users can continue to train DeepTrace models on their own datasets to further improve its performance.

Users train DeepTrace by accessing the DeepTrace Training window in Phoenix by clicking on the train button from the Phoenix launch window.

In the training window, select a single GPU, a model, and a variety of datasets to train with.