An interactive exhibit where you explore and identify the Moon's four major rock-forming minerals under a polarizing microscope, label your own datasets, contribute to a growing dataset for a foundation model of the Moon, and observe how an AI agent working in real time.
Intro
The Moon has no liquid water and no plate tectonics, so its minerals never weathered into clays or micas. As a result, a lunar thin section is a simple mosaic of just a handful of primary minerals, formed billions of years ago and have remained largely unchanged ever since. Two major rock types dominate the lunar surface. The bright highlands are composed mainly of anorthosite, which is almost pure plagioclase feldspar; the dark maria are basalts, containing mostly pyroxene, plagioclase, olivine, and titanium-rich ilmenite. Learn these four, and you can read most of the lunar surface.
Your path through the exhibit
Meet the four minerals.
Examine unknown specimens, record what you observe, and assign a label. You are building a training set.
Turn your labels into a k-nearest-neighbor classifier, score it on fresh slides, then let a language-model petrographer identify a live unknown.
Label minerals using SLIC.
Extract lunar olivine oxide data from Apollo-era PDFs.
Build a lunar sample Agent.
Station 01 — Optical mineralogy
Pick a mineral, and watch how each behaves under crossed polars — that behavior is the diagnostic.
Crossed polars (XPL) · stage rotation 0–360° · grains darken to extinction every 90°
Station 02 — Build the training set
Each slide is an unlabeled specimen. Observe it, record its optical properties, and assign the mineral you think it is. Your judgements become the data the AI learns from — so be honest about what you can actually see.
Record your observations
| id | relief | biref | cleavage | fracture | twinning | opacity | label |
|---|
Aim for at least 2–3 examples of each mineral before moving on. More variety = a better classifier.
Station 03 — From labels to KNN Classifier
Your labeled rows become the memory of a k-nearest-neighbor classifier. To classify an unknown, it finds the most similar slides you labeled and lets them vote. Transparent, no black box.
No model yet — build a dataset in Station 02 first.
The classifier is scored on 40 held-out slides it has never been trained or taught on. The same 40 are reused all session, so you can measure it now, teach it in step 3, then evaluate again — any change in accuracy is real learning, not a different test.
Verdict
Train your classifier, then classify the unknown on the left.
Second opinion — language-model petrographer
Your k-NN agent reasons from numbers. This one reasons from words: describe a thin section and Claude will identify it and explain why. Try the live unknown's description, or write your own.
Station 04 — Label minerals using SLIC
Station 05: Knowledge Base
Station 06: Lunar Sample Agent