Parameters & Run

Input parameter file

The actual input_params.yml from examples/05_westUS_ani:


#######################################
# data parameters
#######################################
data:
  src_rec_file_ph: src_rec_data_wus.csv
  src_rec_file_gr: src_rec_file_gr.csv
  iwave: 2                      # 1 Love, 2 Rayleigh
  vel_type: [True, False]       # phase velocity only
  weights: [0.5, 0.5]
 
output:
  output_path: OUTPUT_FILES_ph_aniso/
  output_initial_model: True
  output_in_process_data: True
  output_in_process_model: True
  log_level: 2
 
domain:
  depth_min_max: [0, 80]        # depth range (km)
  grid_method: 0                # 0: grid by interval
  grid_method_0:
    interval: [0.1, 0.1, 2]     # lon (deg), lat (deg), z (km)
    num_grid_margin: 5
  grid_method_1:
    lon_min_max: [-125.5, -103.5]
    lat_min_max: [29, 49.5]
    n_grid: [221, 201, 41]
 
model:
  init_model_type: 1            # 1: 1-D inversion of average tt
  vel_range: [3.2, 4.8]
  init_model_path: csem.h5
 
topo:
  is_consider_topo: False
  wavelen_factor: 2.5
 
postproc:
  kdensity_coe: 0
  independent_smooth_ani: False
  smooth_method: 0              # 0: Gaussian smoothing
  smooth_method_0:
    sigma: [0.6, 12]            # iso: horiz (deg), vert (km)
    sigma_ani: [0.6, 12]        # aniso: same shape
  smooth_method_1:
    n_inv_components: 5
    n_inv_grid: [20, 25, 20]
    n_inv_grid_ani: [8, 9, 8]
 
inversion:
  is_anisotropy: True           # joint iso + aniso inversion
  use_alpha_beta_rho: false
  rho_scaling: false
  niter: 100
  min_derr: 0.0005
  optim_method: 1               # 1: LBFGS with line search
  step_length: 0.02
  maxshrink: 0.6
  c1: 0.1
  c2: 0.9
  max_sub_niter: 10

Key choices

Parameter	Value	Why it matters
`iwave`	2	Rayleigh wave
`vel_type`	`[True, False]`	Phase-velocity inversion only
`is_anisotropy`	`True`	Joint isotropic + azimuthal-anisotropy inversion
`grid_method`	0 (by interval)	0.1°/0.1°/2 km spacing with a 5-cell margin
`depth_min_max`	`[0, 80]` km	Covers crust and uppermost mantle
`init_model_type`	1	1-D average-tt start — robust without external data
`vel_range`	`[3.2, 4.8]` km/s	Reasonable bounds for western US Rayleigh phase velocities
`smooth_method`	0 (Gaussian)	`sigma = [0.6°, 12 km]` for both iso and aniso
`optim_method`	1 (LBFGS)	Wolfe line search with `c1 = 0.1`, `c2 = 0.9`
`niter`	100	Large iteration budget — anisotropic runs converge slower
`min_derr`	`5 × 10⁻⁴`	Tight stopping criterion

With 0.1° × 0.1° × 2 km and a 5-cell margin, the forward grid is 217 × 206 × 41 ≈ 1.83 million cells. This is a production-scale problem — plan for HPC resources.

Run the inversion

Prepare the run


cd examples/05_westUS_ani
mkdir -p OUTPUT_FILES_ph_aniso

(Optional) Build csem.h5 if you want to use a 3-D initial model — see Data Preparation.

Launch the inversion

Workstation


mpirun -np 16 SURFATT_tomo -i input_params.yml

Monitor convergence


tail -f OUTPUT_FILES_ph_aniso/objective_function.txt

As a reference, on 192 cores (AMD EPYC 9655) 70 iterations complete in about 19 min 44 s (roughly 17 s/iter), so the full 100-iteration run finishes in under 30 minutes. Peak resident memory is ~32 GB (33,535,984 KB). Runtime and memory scale roughly linearly with the number of station pairs, so both may increase significantly on denser datasets.

Expected outputs

After the run, OUTPUT_FILES_ph_aniso/ contains:

File	Description
`final_model.h5`	Final 3-D model (isotropic Vs + anisotropic parameters)
`model_iter.h5`	Model snapshots per iteration
`initial_model.h5`	1-D starting model (when `output_initial_model: True`)
`objective_function.txt`	Misfit vs. iteration
`src_rec_file_forward_PH.csv`	Predicted travel times at the final model