Fault-Detection Training Pairs

Labels for Free

This is where the whole part pays off. A modern fault-detection network learns from thousands of examples, each a seismic section paired with a mask that marks exactly where the fault is. Getting those labels from real data is expensive and subjective, because a human has to pick every fault by hand. A synthetic sidesteps the problem entirely: the fault mask is the fault plane you drew, so the label is exact and free, and you can make as many as you like.

The widget is a working data generator. Each pair is a randomly faulted, folded, layered model pushed through the 2D convolutional section you built in this part, alongside its binary fault mask in red. Regenerate for a fresh batch, and the download is not a picture of a dataset, it is a dataset: a NumPy float32 array of shape (N, 2, height, width) whose first channel is the section and second channel is the mask, ready for numpy.load and a training loop.

Why Convolution Is Exactly Right Here

Recall the fit-for-purpose verdict from Part 2: for fault-detection training data, convolution is not the budget option, it is the correct one. The network learns fault geometry, and geometry is precisely what convolution renders faithfully. Spending a wave-equation solver here would cost a thousand times more compute to add diffractions and multiples that the geometry label does not depend on. Cheap, exact, and unlimited is the whole reason synthetic data and machine learning fit together, and you now have the generator in your hands. The closer of this part steps back to name, honestly, the places where this same 2D convolutional section quietly lies.