An interactive near-Earth-object impact simulator. Enter an asteroid — either as physical properties or as real Keplerian orbital elements — and the app propagates its orbit, decides whether it strikes Earth, works out where on the rotating planet it lands, and models what that does to the ground and the people on it. The result is drawn on a photoreal 3D globe.
Everything runs locally. No API keys, no accounts, no billing relationship with any service.
Real footage of the running app — the approach, the detonation flash, and the damage rings blooming outward. Rendered frame by frame from the WebGL canvas (full-quality MP4).
There is also a 48-second explainer covering the data, the physics validation, and the machine-learning results.
A 900 m stony impactor on the US east coast. Rings run from total destruction (67 km) out to window breakage (2,164 km), drawn as real spherical bands on the globe rather than a flat decal.
The Chicxulub impactor at its historical parameters. The model puts the final crater at 164 km against the ~180 km measured in the rock.
The app is three layers, and each one is checked against something real.
Heliocentric Keplerian propagation in the J2000 ecliptic frame.
- Kepler's equation solved by Newton–Raphson from a Danby starting guess —
residuals stay at
4e-16even ate = 0.995, where the naive guess stalls. - Earth's position from Standish's approximate elements.
- Close approaches found by a coarse sweep, then golden-section refinement.
- Impact decided by gravitational focusing, not by geometric radius:
b_crit = R·√(1 + v_esc²/v_inf²). A slow body is captured from much further out than a fast one. - The landing site comes from solving the hyperbolic encounter about Earth: the incoming asymptote and impact parameter give a hyperbola, we find the true anomaly where its radius equals the atmospheric interface, then rotate that point through the obliquity and GMST into Earth-fixed coordinates.
MOID — the closest the two orbit ellipses ever come, regardless of timing — is computed by a coarse double sweep plus shrinking-window refinement.
| MOID check vs JPL's published values (300 real NEOs) | |
|---|---|
| Median absolute error | 0.000138 AU |
| 90th percentile | 0.000674 AU |
| Phaethon | 0.01870 vs JPL 0.01870 |
| Eros | 0.14864 vs JPL 0.14900 |
Atmospheric entry is a numerical integration of the pancake fragmentation model (Hills & Goda 1993; Chyba et al. 1993) down an exponential atmosphere, rather than the closed-form approximations usually used. The fragment cloud accelerates apart from rest under the ram-pressure gradient — getting this right is what moved Tunguska's burst altitude from 22.7 km to 12.7 km and made Barringer produce a crater at all.
Downstream effects follow Collins, Melosh & Marcus (2005) — crater π-group scaling, seismic magnitude, thermal exposure with horizon clipping, air-blast overpressure, ejecta, and tsunami.
Calibration against real events:
| Event | Quantity | Model | Observed |
|---|---|---|---|
| Chelyabinsk 2013 | energy | 0.52 Mt | ~0.5 Mt |
| Chelyabinsk 2013 | burst altitude | 34.4 km | 29.7 km |
| Tunguska 1908 | energy | 10.8 Mt | 10–15 Mt |
| Tunguska 1908 | burst altitude | 12.7 km | ~8 km |
| Barringer Crater | final diameter | 1234 m | 1200 m |
| Chicxulub | final diameter | 158 km | ~180 km |
That is six orders of magnitude in energy with no per-event tuning.
Two model families, both gradient-boosted histogram trees.
Impact classifier. Features are the six orbital elements plus derived geometry — Tisserand parameter, perihelion/aphelion distance from 1 AU, and crucially the MOID. Labels come from actually propagating each orbit and testing capture.
The interesting property: the impact boundary in element space is razor-thin and chaotic — a metre per second changes the encounter by thousands of kilometres. The classifier cannot memorise it, so its output probability converges on the fraction of nearby orbits that strike, which is exactly the Monte-Carlo impact probability under observational uncertainty.
Adding MOID was decisive:
| without MOID | with MOID | |
|---|---|---|
| ROC-AUC | 0.691 | 0.925 |
| PR-AUC | 0.161 | 0.443 |
| Distance R² | 0.371 | 0.833 |
Trained at full size (110,873 orbits, 8.2% positive) it reaches ROC-AUC 0.944, PR-AUC 0.542, distance R² 0.858 — and it is well calibrated, which is the claim that matters:
| predicted | observed | n |
|---|---|---|
| 0.002 | 0.001 | 15,449 |
| 0.167 | 0.173 | 1,754 |
| 0.373 | 0.362 | 1,537 |
| 0.589 | 0.582 | 1,271 |
A predicted 37% really does strike 36% of the time, so the classifier's output can be read as an impact probability rather than just a ranking score.
Positive examples are ~1-in-10-million under naive sampling, so orbits are drawn from a mixture: most bootstrapped from the real JPL catalogue with noise, the rest constructed backwards from a chosen encounter geometry — pick an epoch, place the body near Earth with a plausible relative velocity, invert the state vector to elements. That lifts the positive class to ~9%.
Effects regressors. Impactor properties to the full damage profile. Two targets are conditional — a body that bursts in the air leaves no crater, one that reaches the ground has no burst altitude — so each is fitted only where it is defined and gated at inference by the airburst classifier.
Trained on 120,000 impactors, every target lands at R² 0.979–0.9999. The UI shows the surrogate and the analytic physics side by side with the error percentage, so the model's accuracy is visible rather than asserted — including the cases where it is poor, such as small iron bodies near the airburst threshold, where the surrogate can be tens of percent out on crater size.
- Two-body Keplerian propagation. No planetary perturbations, no Yarkovsky, no relativistic terms. Real impact monitoring (JPL Sentry) uses full n-body integration; predicted encounter distances here drift from reality over decades. MOID, which is phase-independent, stays accurate.
- Airburst overpressure is evaluated at slant range, which under-predicts mid-field Mach-stem enhancement by up to ~2×.
- Population exposure counts cities over 15,000 people, so rural population is undercounted.
- Tsunami modelling is order-of-magnitude only.
cd orbital-sentinel && ./run.sh --setupOn Windows:
cd orbital-sentinel; .\run.ps1 -SetupThen open http://127.0.0.1:8712. To train the models yourself:
./run.sh --trainTraining takes a few minutes on a laptop CPU. The analytic physics works without the models; only the surrogate columns go blank.
All free, no key, no account.
| Source | Used for |
|---|---|
| JPL Small-Body Database Query API | 42,182 real NEO orbits |
| NASA CNEOS Sentry | 2,184 objects with non-zero computed impact risk |
GeoNames cities15000 |
34,096 cities for population exposure |
| NASA Blue Marble / Earth at Night | globe textures |
Assets are downloaded once and cached, so the app runs fully offline afterwards.
backend/
app/
constants.py physical constants
orbital.py Kepler solvers, ephemeris, encounters, MOID, batch propagation
impact.py pancake entry ODE + Collins-Melosh-Marcus effects
geo.py land mask, city database, population exposure
dataset.py training-set generation
train.py model fitting and evaluation
predict.py inference, Monte Carlo, surrogate-vs-physics comparison
main.py FastAPI service
build_assets.py one-off asset preparation
data/ catalogues and caches
models/ trained artefacts + training report
frontend/
index.html
css/style.css
js/
scene.js renderer, cameras, starfield, bloom
earth.js globe shaders: terminator, city lights, clouds, atmosphere
overlay.js impact marker, damage rings, trajectory, detonation
system.js heliocentric orbit view
ui.js results rendering
main.js orchestration
| Endpoint | Purpose |
|---|---|
GET /api/health |
status, material presets, uncertainty presets |
GET /api/neos?search=&limit= |
real NEOs from the JPL catalogue |
GET /api/sentry |
CNEOS Sentry risk list |
GET /api/geo/terrain?lat=&lon= |
land/ocean and nearest cities |
POST /api/effects |
damage profile only — the slider fast path |
POST /api/simulate/simple |
impactor and impact point stated directly |
POST /api/simulate/elements |
propagate an orbit, resolve any encounter |
GET /api/models |
training report with all metrics |
Collins, G.S., Melosh, H.J., Marcus, R.A. (2005). Earth Impact Effects Program. Meteoritics & Planetary Science 40, 817–840.
Hills, J.G., Goda, M.P. (1993). The fragmentation of small asteroids in the atmosphere. Astronomical Journal 105, 1114.
Chyba, C.F., Thomas, P.J., Zahnle, K.J. (1993). The 1908 Tunguska explosion. Nature 361, 40–44.
Standish, E.M. (1992). Keplerian Elements for Approximate Positions of the Major Planets. JPL Solar System Dynamics.


