IBM x Airbus Hackathon on wing corrosion dataset ✈
https://www.kaggle.com/competitions/haks-airbus-x-ibm-x-aws-2026
Airbus aircraft are corroding much faster than expected, causing significant dissatisfaction among airlines. Aircraft undergo scheduled maintenance visits ("C-checks", e.g., after 6 years) where parts are repaired or replaced — but corrosion problems are appearing before these scheduled visits, meaning airlines discover issues during operations. Severe corrosion on wings is extremely costly to repair, requiring corroded sections to be cut out and replaced with plates.
A striking real-world example: during COVID, parked aircraft corroded so badly that some landing gear collapsed. Airbus discovered that corrosivity varies enormously between airports — some (e.g., desert locations) are 10x less corrosive than average, others 6x more, meaning a factor of ~60 between the least and most corrosive airports. By providing airlines with a corrosivity map so they could park aircraft at less corrosive airports (allowing mandatory $50k inspections to be stretched from every 12 weeks to every 24 weeks), Airbus saved airlines roughly €400 million — a concrete proof that predicting corrosion creates value.
The goal is essentially a "digital twin" of wing degradation: estimating the state of corrosion of an aircraft's wings at any point in time.
Concretely, you must add a column to the test files containing, for each aircraft and each month, the probability (0–100%) that corrosion would be found if the aircraft were inspected at that time.
- Training data: aircraft from 2014 onwards, with both their monthly environmental data and observed corrosion events.
- Test data: only aircraft delivered in 2014 (all "born" in 2014), with environmental data but without the corrosion observation file.
Important domain knowledge: corrosion is non-linear — once it starts, it progresses exponentially fast.
All data is anonymized (aircraft identities belong to airlines, not Airbus). The sources are:
-
Corrosion events from two channels:
- Logbooks: maintenance crews walk around the aircraft, note corrosion in the aircraft's logbook. These belong to airlines; some share them, others don't. Most corrosion events come from here.
- Tech requests: when an airline finds corrosion, they contact Airbus support (often in a panic) asking what to do — these requests also flag corrosion events.
-
Environmental data (Copernicus): via Airbus Defence and Space, the European Copernicus satellite program provides global air-quality data on a ~7 km grid, with about 50 "contaminants" (salt, pollution, etc.).
-
METAR data: meteorological sensors located at airports, providing humidity, wind, temperature, visibility — very precise since the sensors are on-site.
-
Flight Radar data: purchased from a private company, refreshed roughly every minute, with history back to 2013, used to know where and when each aircraft is on the ground.
Key insight: aircraft spend roughly two-thirds of their time parked on the ground (especially short-haul). In flight, all aircraft share essentially the same environment; it's on the ground that environments differ (near the sea, polluted areas, etc.). So Airbus computes the environment only for ground time (stays of at least ~20 minutes), and you'll receive, per aircraft, the monthly average ground environment plus the duration spent on the ground that month (some aircraft are parked the whole month, others fly a lot).
The metric is the Brier score — lowest score wins. Evaluation works as follows: Airbus knows the actual corrosion observation dates for the test aircraft (~150 aircraft, two evaluation dates each, ~300 samples). For each aircraft they check your predicted probability at:
- The corrosion observation date → expected to be close to 100%
- Two years before that date → expected to be close to 0% (since corrosion develops quickly, it wasn't present 2 years earlier)
You're not forced to output only 0s and 100s — intermediate probabilities are allowed and often wiser, because a confident wrong answer (predicting 0 when 100 was expected) is heavily penalized by the Brier score. Calibrating your risk estimates is part of the challenge.
You'll also present a pitch explaining your approach. The presenter cares not only about methodology but about what you learned about corrosion and how you explain your results — he explicitly hopes to be surprised and to learn something Airbus hasn't discovered yet.
Airlines initially rejected the idea of a corrosion prediction model ("I don't want predictions, I want no corrosion — a prediction just forces costly maintenance on me"). Airbus's answer leveraged the fact that aircraft fly repetitive route patterns, so their environment is roughly stable over their lifetime. The proposed solution: swap aircraft between routes so that fast-corroding aircraft are moved to gentler environments, ideally ensuring all aircraft reach corrosion only at end-of-life. The takeaway the presenter insists on: prediction itself solves problems and makes money — "knowledge has value."
Dataset is Apache V2.0