Example: Multiscale Inversion with Random Source Encoding

Script: examples/example_multiscale_random_sources.py

Goal

Show how random source encoding can reduce per-iteration cost while preserving multiscale inversion behavior.

Inputs

• Base model tensors for epsilon, sigma, and mu
• Source wavelet bank and random encoding vectors
• Source and receiver geometry per shot
• Inversion schedule settings (bands, iterations, optimizer)

Steps

1. Build observed data or load precomputed traces.
2. Encode multiple physical shots into randomized super-shots.
3. Run staged inversion from low to higher effective frequencies.
4. Update model parameters using configured optimizer.
5. Track loss and model snapshots per stage.

Outputs

• Loss curves over iterations/stages
• Intermediate and final epsilon reconstructions
• Optional comparisons between encoded and reference traces