On 12 August 2026 the Moon removes the Sun for 37 seconds over Colmenar Viejo, north of Madrid. Today the Sun’s magnetic field has already been measured. This is an attempt to cross the gap between those two facts, and then find out — by looking up — how badly it went.
A potential-field reconstruction of the coronal magnetic field, a topology-informed electron-density model, and an exact Thomson-scattering renderer. Not a thermodynamic MHD model, and it does not pretend to be one. Every approximation is written down and ranked by risk.
// field guide
The Sun will be very low: 7.4° altitude at azimuth 283.2°, toward west-northwest (WNW). A clear horizon matters more than gaining a few metres of proximity: buildings, trees and the Guadarrama range can hide totality.
Use trustworthy ISO 12312-2 solar viewers throughout every partial phase. Remove them only after ALL of the photosphere has disappeared; replace them as soon as the first bright point appears and, if in any doubt, before C3. Ordinary sunglasses and homemade filters are not safe.
Cameras, binoculars and telescopes need a purpose-made solar filter secured in front of the objective. Eclipse glasses behind concentrating optics are dangerous. If you are improvising, do not handle magnifying optics during transitions: observe unaided with the correct viewer.
| First contact (C1) | 19:36:24 |
|---|---|
| Totality begins (C2) | 20:31:42 |
| Maximum | 20:32:00 |
| Totality ends (C3) | 20:32:24 |
| Sunset | 21:17:06 |
Totality lasts about 37–42 seconds, depending on the exact location and calculation model. It is not universal across Colmenar Viejo, and a clock never replaces checking that the photosphere has disappeared.
Keep viewers on throughout partial phases and observe ambient light, temperature, the horizon and animal behaviour. A cloud is never a solar filter, even when the Sun looks dim.
During partial phases, gaps between leaves act as pinhole cameras and project crescents; shadows sharpen and the light loses saturation, sometimes appearing metallic. Near totality, shadow bands may appear: moving patterns produced by atmospheric turbulence, subtle and not guaranteed.
At the contacts, lunar valleys pass points of photospheric light: Baily's beads. The last or first bright point beside the corona forms the diamond ring. The chromosphere and prominences look red because of strong hydrogen emission; the lunar disk looks black because its night side blocks a vastly brighter photosphere.
The corona looks white because free electrons Thomson-scatter the full mixture of photospheric light without selecting one dominant visible colour. The sky does not become full night: outside the narrow umbra, the atmosphere and peripheral horizon remain illuminated, producing 360° twilight. The umbra may be visible approaching or departing across the landscape.
At only 7.4°, light crosses much more atmosphere: aerosols, humidity and haze increase extinction and reduce contrast and colour, making the WNW horizon critical. Venus is plausible, but do not sacrifice the corona searching for it. Cooling, wind and animal responses vary greatly and may be confused with sunset; they are worthwhile observations, not guarantees.
// method
The corona is visible during totality because free electrons scatter photospheric light. Its shape is set almost entirely by the magnetic field. So the chain runs from a measured surface field to a synthetic image, and every link introduces error worth naming.
what this is not
Predictive Science run a time-dependent MHD model for this eclipse, assimilating far-side magnetograms from Solar Orbiter/PHI. They will be better than this. That is the point of having them as a baseline — the interesting question is where a cheap model breaks, and by how much.
// phase A — complete
The dominant error in this whole pipeline is not the renderer and not even PFSS. It is that we cannot see the far side of the Sun. Global maps reconstruct the hidden hemisphere with a surface flux-transport model, and August 2026 is near solar maximum, when active regions evolve in days.
ADAPT ships twelve realisations of that reconstruction. They agree almost exactly on how much flux exists — total unsigned flux varies by only 1.14%. They disagree sharply on where it is, and the disagreement is organised by longitude.
| Region | Carrington longitude | Spread |
|---|---|---|
| Disk centre | 224.5° | 1.85 G |
| West limb | 314.5° | 4.11 G |
| East limb | 134.5° | 7.71 G |
The east limb is the edge that most recently rotated out of the far side, so it is the longest unobserved. It carries 4.2× the uncertainty of disk centre — and on an eclipse image the limbs are exactly where the structure lives, because streamers are seen edge-on against the sky.
a pre-registered prediction about our own failure
The preregistered prediction states that, if this model fails, it should fail asymmetrically, worse on the east limb. If it fails symmetrically, the far-side field was not the limiting factor and something else is — the source surface, or the density proxy. Either way we learn something, which is more than a confident picture would have given us.
// phase B — complete
PFSS assumes the corona carries no currents, so the field is the gradient of a potential satisfying Laplace’s equation. The model’s real physical content is its upper boundary: a source surface at radius Rss where the field is forced radial, standing in for the solar wind dragging it open.
There is no such surface. Rss = 2.5 R☉ is a convention from 1969, and recent work benchmarking against eclipse images finds the best value moves with the solar cycle. So we do not pick one — we run five, across all twelve realisations. Sixty solves, 124 seconds on a laptop.
Before trusting any of it: for a pure dipole boundary the potential has a closed form, and the solver must reproduce it.
Br(r, θ) = b₀ cosθ · (Rss⁻³ + 2r⁻³) / (2 + Rss⁻³)
It does, to 5% at the source surface, with the polarity the right way round, and it reproduces the input boundary map to 0.03% RMS. Sign conventions, normalisation and the upper boundary condition are all pinned by that single test.
// phase C — complete
With the field solved, field lines are integrated from 16,200 equal-area photospheric seeds and classified. Closed lines return to the surface at both ends: they trap plasma, are overdense, and build the streamers. Open lines reach the source surface: plasma escapes as solar wind, the region is underdense, and it forms a coronal hole — dark in white light.
That classification is what turns a magnetic field into something we can put a density on, so it feeds directly into Phase D.
The classification is converged: doubling the integration step budget changes zero seeds. That matters, because a line that exhausts its budget mid-flight is indistinguishable from one that escaped.
| Rss | Open area | Boundary spread |
|---|---|---|
| 1.3 R☉ | 30.69% | 1.64 pp |
| 1.5 R☉ | 20.43% | 1.10 pp |
| 2.0 R☉ | 11.26% | 0.72 pp |
| 2.5 R☉ | 7.74% | 0.70 pp |
| 3.0 R☉ | 5.99% | 0.56 pp |
this ratio is a clean one
Open area varies 26× more across source surfaces (24.7 percentage points) than across boundary realisations (0.94 pp). Unlike open flux, nothing here is definitional: open area is always measured at the photosphere, a fixed radius, while the parameter being varied lives far above it. Choosing Rss = 2.5 instead of 1.5 changes how much of the Sun is open from 20% to 8%.
// phases D and E — complete
This is what the model, written before totality, predicted for a field north of Madrid on 12 August 2026 at 20:31 CEST. Zenith is up, so this is the orientation the eye sees, not the solar-north-up convention: solar north is tilted 34.7° from vertical.
The Sun is not a point source — an electron at 1.5 R☉ sees a disk 42° wide — so the scattering angle varies across the disk and must be integrated over it with limb darkening. Van de Hulst did that integral in 1950, leaving four closed-form coefficients. The decisive test is that far from the Sun the full finite-disk kernel must collapse onto the textbook dipole pattern:
B(χ) / B(90°) → 1 + cos²(χ) as r → ∞
It does, to 0.2%. That single check pins the sign, the normalisation and the whole disk integral at once — nothing else in the kernel can be wrong while it passes.
| Kernel evaluations | 415M per frame |
|---|---|
| Render time | 118 s (NumPy, CPU) |
| Field lines traced | 884k |
| Quadrature convergence | 0.98% worst case |
| Peak memory | 1.8 GB |
what is honest here and what is not
The geometry is real: observer position, solar orientation, scattering angles, the occulting disk. The scattering is exact. The magnetic topology is a genuine PFSS solution from a real magnetogram.
The density is a proxy — closed field 3.5× enhanced, open field 0.4× depleted, over a Baumbach–Allen radial profile. Those are knobs, fixed before any comparison with Predictive Science so we cannot tune our way into agreement.
And the structure is too smooth. Real coronae show fine radial striations; a potential field with a two-valued density proxy cannot produce them. If the observed corona is sharper and more filamentary than this, that is the density model failing, not the renderer.
// simulation
These are not the same picture at different scales. As Rss falls, the belt stops being a single band and breaks into several distinct streamers. These predict visibly different eclipses.
ADAPT-GONG 2026-08-11T04:00Z · realisation 0 · nr = 100
Open flux varies 48× more across the five source surfaces than across the twelve boundary realisations. We are not going to lead with that number: open flux is measured at the source surface, whose radius is the parameter being varied, so part of that ratio is definitional rather than physical.
The honest metric is the one an eclipse actually shows: where the streamer belt sits on the sky.
| Across 12 ADAPT realisations, fixed Rss | 5.3° |
|---|---|
| Across 5 source surfaces, fixed realisation | 20.5° |
| Ratio | 3.8× |
So the ranking survives with a corrected magnitude. The source-surface choice moves the streamer belt about four times as much as the entire far-side uncertainty, not fifty. Both matter; the fictional parameter matters more.
// post-eclipse check
provisional · qualitative only
Six frames from the official ESA/Javalambre telescope feed, between 01:03:50 and 01:04:40, consistently show a non-circular, multipolar corona with roughly 3–4 broad fans and a visual extent of ~2–2.5 R☉. That coarse morphology is compatible with the frozen prediction, but it is not an official score.
The available feed is compressed 1920×1080 video and does not
document solar north, optical parity or calibration. It therefore
cannot support position-angle (PA) matching, test the preregistered
east-west asymmetry, or rigorously measure radial extent. Its status is
qualitative_only; the official score remains pending.
| Source | Available provenance | Official-score blocker |
|---|---|---|
| Commons · Logroño | 3000×4000 JPEG; EXIF 18:28:20 UTC; GPS; CC BY-SA 3.0. | CustomRendered; optical chain, parity and coverage to 2.5 R☉ cannot be certified. |
| IGN/OAN · Yebes | Institutional Solar Tower, white light. | Stream without RAW/FITS, orientation or calibration. |
| IAC/NATE | Institutional observing programme. | No suitable public originals yet. |
official_candidate: none_available
The official validation protocol requires selecting an observation by provenance before comparing its morphology. Waiting for better data protects the result from choosing, after inspection, whichever image most favours the prediction.
The qualitative check used six frames distributed from 01:03:50 to 01:04:40 in the linked feed. No third-party images are embedded, and these frames were not used to run the official protocol.
the prediction remains prior
The prediction and its source magnetogram were published before totality, with the input observation time recorded. This later check does not modify those artefacts or their history.
