Hyperspectral imaging sees the whole rock face but only infers grade. LIBS measures grade, one spot at a time. The Vulcanis loop uses the HSI scan to decide where each LIBS shot will reduce map error the most.
Play the loop, then switch strategy to compare it with random, grid and spectral-novelty sampling on the same scene and the same shot budget.
Simulated SWIR scene with known ground truth. An algorithmic benchmark, not field data.
HSI cue SWIR band-depth composite
R 2200 nm (sericite, kaolinite) · G 2340 nm (calcite) · B 2170 nm (alunite)
True grade hidden from the loop
01.2 % Cu
Estimated grade after 0 shots
01.2 % Cu
Uncertainty posterior std. dev.
lowhigh
Computing baselines…
Map error (RMSE)–
Ore/waste misclassified at 0.5 % Cu–share of the face
Map error vs. shots fired
Same scene, same model, same LIBS noise; only the choice of shots differs. Random is the mean of 10 runs.
How the simulation works, and what it does not prove
The loop, one shot at a time
A 72 × 72 pixel SWIR cube (2000–2490 nm, 50 bands) is synthesised from five mineral endmembers with realistic absorption positions, plus topographic shading and sensor noise.
The cube is unmixed into mineral abundances. These are the cue: they correlate with grade, but part of the grade is controlled by factors HSI cannot see.
A Gaussian-process posterior combines the HSI abundances with spatial correlation between shots.
The loop fires where the posterior is least certain, reads the LIBS result (true grade plus measurement noise), updates the map and chooses again.
Baselines
Random: shots spread at random over the face.
Systematic grid: an evenly spread low-discrepancy pattern, the usual manual sampling logic.
Spectral novelty: fires at the most spectrally unusual pixel not yet covered, in the spirit of AEGIS-style autonomous targeting on Mars rovers.
All four feed the same posterior model, so differences come only from where the shots land. On some scenes a baseline ties or wins at some budgets; those cases are part of the failure envelope, not hidden.
Limits
Scene, endmembers and grade relationship are synthetic; ground truth is known by construction.
LIBS is modelled as grade plus Gaussian noise: no plasma physics, matrix effects or standoff attenuation.
Model hyperparameters are fixed, not learned per scene.
The geologist-picked baseline is not yet included.
This validates the decision logic in simulation (Phase 0 of the TRL roadmap). Next steps: replace the synthetic cube with open core-scan or outcrop cubes, then replace the LIBS stand-in with measured spectra.