Example: Multiscale Filtered Inversion

Script: examples/example_multiscale_filtered.py

Goal

Demonstrate multiscale inversion by progressively increasing data bandwidth using FIR low-pass schedules.

Inputs

• Model file: examples/data/OverThrust.npy
• Key parameters: dx, dt, nt, pml_width, n_shots, storage_mode

Steps

1. Generate base observed data at a fixed forward frequency.
2. Apply FIR low-pass filters to create multiscale datasets.
3. Run staged inversion (AdamW then LBFGS per stage).

Outputs

• Filtered data comparison image.
• Stage snapshots of epsilon.
• Summary plot with loss curve.

Notes

• CUDA is strongly recommended for practical runtime.
• Runtime depends on grid size, shot count, and inversion stage count.
• For debugging, reduce nt and n_shots first.