Wreckage localisation methodology

This document describes the model used by SeyAero to estimate the location of an aircraft following a hull loss, and to plan the search for it. It is written so that a technical reviewer in an investigating authority, a search coordinator, or an insurer's engineering assessor can judge the model on its assumptions rather than on its output.

The model produces a probability distribution over position. It does not determine cause, apportion blame, or replace the judgement of the investigator-in-charge. Every output is decision support for search planning under ICAO Annex 12 and Annex 13 processes.

The model consumes only evidence held in the case file:

• Last known flight state — position, altitude, ground speed, track, vertical speed, and a fix confidence in [0,1].
• Full position track where available — radar, ADS-B, ACARS or satellite handshakes — used for the multi-fix trajectory fit.
• Meteorology by altitude band — wind direction and speed, temperature, visibility, phenomena.
• Oceanography — surface current set and drift rate, sea state, and water depth from the GEBCO 2020 grid.
• Ground and witness reports — observer position, bearing, distance, and an investigator-assigned reliability in [0,1].
• Negative search effort — swept areas with an assessed probability of detection.

Meteorological and oceanographic fields can be pulled automatically from ERA5 reanalysis and marine reanalysis for the occurrence position and time, rather than entered by hand.

Where more than two fixes exist, a segment-wise fit derives track angle, turn rate, vertical speed and specific-energy trend across the final portion of the track. Specific energy is computed as altitude plus the kinetic term v²/2g expressed in feet, which exposes energy bleed that a raw altitude trace hides.

The residual scatter of segment bearings about the fitted turn gives a 1σ heading uncertainty. That uncertainty is carried into the Monte Carlo as an additional dispersion term rather than being discarded, so a noisy track widens the answer instead of producing a falsely confident one.

Between 500 and 20,000 particles are released from the last known state. Each particle is advanced through four stages.

• Glide footprint. A lift-to-drag ratio is sampled uniformly: 10–18 for a controlled glide, 0.8–4.5 for an out-of-control descent. Forward travel is (altitude / 6076.12) × L/D nautical miles.
• Departure heading. Dispersion about the last known track scales inversely with the sampled glide ratio: a steep, low-L/D descent has ground travel small relative to height lost, so the departure direction is only weakly constrained by the last recorded track. Dispersion runs from roughly 18° for a controlled glide to 120° for a near-vertical departure.
• Wind transport. Descent duration is altitude divided by the sampled descent rate; mean wind for the release altitude band advects the particle over that duration, with 25% speed variance and 12° directional variance.
• Water-column behaviour. Heavy structure sinks at 1.0–2.2 m/s and is displaced by current only for the sink duration implied by local depth, then scattered by break-up. Buoyant debris is advected by surface current for the full elapsed window, plus leeway at 2–4% of surface wind, plus turbulent diffusion growing as √t.

Impact and floating populations are solved as separate fields. The impact field is the sonar and seabed search zone; the floating field is the aerial and surface search zone. Conflating the two is a common cause of wasted search effort in real campaigns.

Ground and witness reports act as soft anchors. Each is converted to an implied position — directly, or by projecting bearing and distance from the observer — with a positional σ of 12–50 km scaled by the investigator's reliability rating. Particle weights are multiplied by a Gaussian likelihood raised to the power of the reliability, floored so that no single report can veto a particle outright. An unreliable report nudges the field; it cannot move it.

Negative search effort is applied with Koopman's law. A particle inside a swept area survives with probability (1 − POD) per pass. The field is then renormalised, so probability flows out of searched water and into unsearched water. This is why each fruitless sweep sharpens the next one rather than merely crossing an area off.

Sweep planning uses the standard maritime SAR random-search law. For an asset of speed V and effective sweep width W spending time t on a sector of area A, coverage is C = V·W·t/A and probability of detection is 1 − e^(−C).

Effort is allocated greedily by marginal probability of success per hour. For exponential detection functions this greedy allocation is the optimal allocation, so the resulting plan is not merely a heuristic ordering — it maximises the chance of finding the wreckage in the hours available.

The engine is a pure function of its inputs and a seed. Re-running a saved solution with its stored parameters and seed reproduces it bit for bit, which is what makes the output admissible as a documented analytical step rather than an opaque assertion.

Every action on a case is written to an append-only activity log. Each entry is hashed together with the hash of the entry before it, forming a chain: altering, deleting or re-ordering any historical entry invalidates every hash after it, and verification reports the exact sequence number where the break occurs. When a report is issued its fingerprint is bound to the chain head at that moment, so a report can be proven to correspond to a specific, unmodified investigative history.

The model is back-tested against real hull losses where the wreckage was subsequently recovered. Each case is re-run using only the information searchers held at the time; the recovery position is withheld from the model and used solely for scoring. Results, including any misses, are published in full and are stable across seeds.

These are the conditions under which the model should be treated with caution:

• Single-fix priors. With one position and no track, heading is unconstrained and the field is correspondingly broad. The result is honest but wide.
• Long unpowered flight after the last fix. Where an aircraft continued for tens of minutes beyond its final fix, the dominant uncertainty is duration, not aerodynamics, and the model's dispersion is optimistic.
• Coarse ocean fields. Reanalysis currents resolve to roughly 0.08°, which under-resolves eddies that matter for day-one debris drift.
• Witness bearing bias. Human bearing estimates are systematically biased and night-time distance estimates are unreliable; reliability ratings should reflect that.
• Seabed relief. High relief creates sonar shadow zones, so an assessed probability of detection over rough terrain should be reduced accordingly.
• No terrain masking on land. Over land the model does not currently account for slope, canopy or terrain channelling of the descent.

The model is calibrated for transport-category fixed-wing aircraft. Rotorcraft, light aircraft and unmanned systems have different descent behaviour and are outside the validated envelope.

Model changes are versioned and re-validated against the full benchmark set before release; a change that improves one case but degrades containment across the set is rejected. Benchmark reference positions are taken from published official reports and cited on the validation page.

SeyAero outputs are advisory. The investigating authority remains the sole determiner of findings and causes under ICAO Annex 13, Chapter 5.