FLARE: Fast Lookup-based Acoustic-seismic Reconstruction of Entries

The fields that matter for a new event are event, paths, picks.file, and the mcmc.parameters bounds. Resolution knobs. Lookup-table cost scales as n_levels x (az_range / az_step) x (incl_range / incl_step): yaml: lookup: altitude_min_km: 15 altitude_max_km: 100 altitude_step_km: 1.0 # altitude level separation azimuth: {min: -180, max: 180, step: 1.0} inclination: {min: -90, max: 90, step: 1.0} Sub-kilometre altitude steps work: the levels are discovered by scanning lookup_npy/, and the forward model brackets each trajectory sample between whichever two levels exist.

A finer angular step lets you tighten forward.knn.max_distance_deg, which is the main lever on interpolation error in the arrival times. Parameter bounds are written in physical units. `log10: true` makes the sampler work in the logarithm while the bounds stay readable: yaml: ablation: {min: 2.5e-9, max: 5.0e-8, log10: true} mass: {min: 300.0, max: 2500.0, log10: true} Order in mcmc.start.guess is always: lon, lat, height, azimuth, inclination, velocity, time_shift, ablation, spherical_factor, mass. ## Running the MCMC mcmc.pool selects mpi (schwimmbad MPIPool, for mpirun), multiprocessing, or serial.

With h5py installed the chain is written incrementally to mcmc/mcmc.h5 and re-running the same command resumes where it stopped. Delete the .h5 to start fresh. Convergence: the run stops once `total_steps >= target_tau_multiple

• max(tau)` with a finite autocorrelation estimate in every dimension, or at `max_steps`. Walkers whose acceptance fraction falls below `reseed_threshold` are reseeded from the healthy ones. ## Post-processing bash: FLARE postprocess -c koblenz.yaml --burn-in 0.3 --draws 500 --burn-in takes a fraction (< 1) or a step count. Besides the posterior table it writes the MAP trajectory as CSV and an altitude-binned lat/lon spread over --draws posterior samples and both are ready to feed straight into PyGMT. mcmc_residuals.csv gives observed, predicted and residual per station at the MAP solution, which is usually the fastest way to spot a mis-picked station. Corner and trace plots need matplotlib and corner. Without them the CSVs are still written. ## Using the pieces directly The stages are library code, not just CLI: python: from meteor_pipeline.config import load_config from meteor_pipeline.context import ForwardContext from meteor_pipeline.forward import forward_model cfg = load_config("koblenz.yaml") ctx = ForwardContext.from_config(cfg) # atmosphere + lookup + picks pred = forward_model(ctx, lon=5.5, lat=49.2, height_km=85.0, azimuth_deg=50.0, inclination_deg=-15.0, velocity_ms=18000.0, time_shift_s=0.0, ablation=1.7e-8, spherical_factor=1.0, mass_kg=1000.0) ## Tests bash: python tests/test_smoke.py # or: python -m pytest tests -q The smoke test builds a toy atmosphere and a synthetic straight-ray lookup table, then runs conversion, the forward model, the posterior and a short serial chain. It needs no infraGA, no network, and takes a few seconds.