42 QuakeHeroes

Problem

Planetary landers cannot afford the power to send continuous seismic data home, yet only a fraction of it contains quakes.

Solution

A pipeline that band-pass filters Apollo and Mars InSight ground-velocity data, decomposes it with empirical mode decomposition, computes sliding-window energy per mode into a 2D matrix, and trains a U-Net encoder-decoder against Gaussian labels centred on each event's arrival time. A companion web page shows detected moonquake and marsquake types.

Demo moment

Not described in the sources.

Why it stood out

• organizer statedWon the award for the most innovative use of technology.Source ↗
• organizer statedUsed a deep neural network to locate seismic events precisely in time-series data, with advanced signal processing to isolate components of non-stationary signals.Source ↗
• team statedValidated on held-out data: 63 lunar events for training, 13 for validation, and both available Mars cases used only to check that a Moon-trained model generalizes.Source ↗
• curator inferenceThe write-up explains each step and why (filtering alone failed on low signal-to-noise, so they added decomposition), which makes the depth visible to expert judges.

Lessons for your next hackathon

• For a technology award, show the pipeline and the reason for each step, not only the result.
• State the validation split, especially when the target environment has almost no data.
• LATAM precedent: a team from Brazil won Best Use of Technology.