Debugging complexity maps#
So far, we have come across two common issues that may arise when doing wavefield learning and we have assembled a (hopefully exhaustive) list of potential causes.
Problem 1: Nr is 1 in parts of or the entire complexity map#
This may be caused, either by a mistake in the 3D model input, or by a source with such low amplitude, that the arrivals do not pass the threshold of numerical error. If you experience this issue, try the following:
Check if the 3D models are loaded properly (for a homogeneous model Nr = 1 would be correct).
Check if the wavefield reaches all parts of the modeled domain within the simulated time
Try to decrease absolute_amplitude_skipped or increase the source amplitude
Problem 2: Nr is maxed out in seemingly random parts of the map and the rest looks fine#
More often than not, raising the starting Nr cap fixes the problem. To make a better diagnosis, guidelines are given here. These should be your next steps:
Raise the Nr cap (but do not disable bound_Nr_by_inplane)
If the problem persists, experiment with increasing relative_amplitude_skipped and absolute_amplitude_skipped, and potentially lowering threshold_Fourier_convergence
Setting max_num_peaks: 1 results in an Nr map which does not account for multiple reflections, but it is the most robust method with respect to noise and hence good for debugging