# Team Crater2Impactor

- **Event:** [Meta Global AI Developer Hackathon](https://cerebralvalley.ai/e/meta-ai-developer-hackathon)
- **When:** Oct 26, 12:00 PM – Nov 4, 9:00 PM (EST)
- **Where:** Online
- **Team:** [Adarsh Pandey](https://cerebralvalley.ai/u/adarshpandey)
- **GitHub:** https://github.com/Adarsh-61/Crater2Impactor
- **Demo video:** https://youtu.be/xGizPZZZTlc
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5222

Crater2Impactor: Impactor Parameter Estimation from Lunar Impact Craters.

The Moon keeps a record of impacts, but a crater does not identify the body that made it: the same 30 m crater fits about 556 kg at 12 km/s or about 130 kg at 37 km/s. Single "best estimate" calculators hide this. Crater2Impactor returns a full probability distribution for the impactor's mass, energy, momentum, size, speed and angle from a measured crater, and labels each quantity as constrained by the crater, set by the prior, or dependent on assumptions.

How it works: a Bayesian engine combines published crater-scaling physics (three named constant sets), the LROC lunar mare map at the crater's location, and declared impactor populations. It solves the scaling law exactly for each draw of the unknowns (importance sampling, 30,000 draws per scenario), adds optional independent measurements (speed, angle, flash energy, ejecta pattern), and reports reliability checks with every answer. We also trained OT-FMPE, a flow-matching neural posterior estimator (1.4 M parameters, 2.4 M simulated craters, 2 x T4 GPUs), and compared the two on 1,114 shared test cases.

Results (verified, reproducible): inside its training range the network ties the engine (mass error 0.47 dex, 90 % coverage 0.90); outside it, it fails without warning (3.26 dex vs 0.76 dex). On 11 real spacecraft impacts with known mass, all 11 true masses fall inside the engine's 90 % range. Limitations are reported openly: no natural impactor has a known mass, and the angle-from-ejecta model failed its one independent check.

Pre-existing work and hackathon work: the project began on 2 Sep 2026. During the hackathon window (26 Oct - 4 Nov) we redesigned the mass prior, built the deterministic quadrature backend, retrained and benchmarked OT-FMPE, released versions 2.0 and 2.1, ran two full code audits with independent verification, and prepared the open-source release and demo.

## More from Meta Global AI Developer Hackathon

- [sanskar](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5219)
- [Cheese Lord](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5220)
- [Oceanio](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5221)
- [team_id=1114754401265198](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5223)
- [Team Techno](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5224)
- [RoboCoach: Your AI Tutor for Working With Robots](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5225)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5222. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt