an interactive field guide

Searching for
Exploding Black Holes

Half a century ago Stephen Hawking predicted that black holes evaporate — and that the smallest ones end their lives in an explosion brighter than anything gravity has built since. Somewhere in the data of our γ-ray telescopes, that flash may already be waiting.

begin ↓
01the physics of dying black holes

Black holes are not quite black

In 1974, Hawking showed that quantum effects near the event horizon force black holes to radiate like thermal bodies . The temperature is set by nothing but the mass:

kBTBH=c38πGM    (1016gM)MeVk_B T_{\mathrm{BH}} = \frac{\hbar c^3}{8\pi G M} \;\approx\; \left(\frac{10^{16}\,\mathrm{g}}{M}\right)\,\mathrm{MeV}

The rule is merciless: smaller means hotter. A stellar-mass black hole radiates at a hundred-millionth of a kelvin — utterly invisible. But shrink one to the mass of a mountain and it glows in γ-rays. And because radiating mass makes it smaller still, evaporation runs away. The final moments are an explosion.

dial a black hole
Black hole mass5×10¹⁴ g
Hawking temperature
21.1 MeV
k_B T = ħc³ / 8πGM
Remaining lifetime
1.6×10⁹ yr
longer than human history
Best current detector
Fermi GBM
peak emission near 90.8 MeV

scale: ≈ a large mountain — evaporating today← M_U = 5.1×10¹⁴ g: lifetime = age of the universe

Lifetime from τ ≈ 407 s × (M/10¹⁰ g)³ — Boluna et al. Eq. (4.4). The 'best detector' targets the band where emission peaks (≈ 4.3 k_BT for a photon blackbody-like spectrum).

What does the light of a dying black hole look like? Two components : direct photons radiated straight off the horizon, and a fragmentation component — quarks and gluons that hadronize into pions, whose decays flood the spectrum with γ-rays. For any hole hotter than the QCD scale, fragmentation dominates.

the photon spectrum, mass by mass
mass10¹⁴ g
10⁻⁷10⁻⁵10⁻³10⁻¹1010³10⁵10⁻²10³10⁸10¹³10¹⁸10²³10²⁸photon energy E [GeV]d²N/dE dt [GeV⁻¹ s⁻¹]k_BT = 106 MeV
totaldirect (primary Hawking)fragmentation (quark/gluon jets → π⁰ → γγ)

T = 106 MeV · remaining lifetime ≈ 1.3×10⁷ yr. Parameterization from Ukwatta et al. (2016) Eqs. 31–34, as implemented in Analytical_Modelling.ipynb.

Grid computed by the site's data pipeline with the exact parameterization used in Analytical_Modelling.ipynb (Ukwatta et al. 2016, Eqs. 31–34), the same forms fit to BlackHawk output.

The mass-loss rate integrates all of this over every available species:

dMdt=α(M)M2τBHM33α\frac{dM}{dt} = -\frac{\alpha(M)}{M^2} \quad\Rightarrow\quad \tau_{\mathrm{BH}} \sim \frac{M^3}{3\alpha}

This produces the strangest lightcurve in astrophysics. Every other transient rises and fades. A black hole explosion only ever brightens, following a universal curve fixed by known physics — right up to the instant the hole ceases to exist.

the backwards burst
10⁻⁹10⁻⁷10⁻⁵10⁻³10⁻¹1010³10²⁷10²⁹10³¹10³³10³⁵1 / remaining lifetime [s⁻¹] → explosionphoton output [s⁻¹]
scrub
remaining lifetime · 16.5 hr
mass left · 5.3×10¹⁰ g
output · 3.6×10³⁰ γ/s
universal lightcurve — it only ever brightens
Total photon output per second versus (inverse) remaining lifetime. Scrub toward the explosion: in the last second, output climbs five orders of magnitude.
02where small black holes come from

Forged in the first second

No star can make a small black hole — stellar collapse bottoms out near three solar masses. But the infant universe could. In its first fraction of a second, space was filled with dense plasma; any patch with density contrast δ0.45\delta \gtrsim 0.45 would collapse the moment it entered the causal horizon . The mass swallowed is the mass inside the horizon at that instant:

MH(t)1015g×(t1023s)M_H(t) \approx 10^{15}\,\mathrm{g} \times \left(\frac{t}{10^{-23}\,\mathrm{s}}\right)

So the cosmic clock doubles as a mass dial: when a primordial black hole (PBH) formed sets how big it was born — and therefore when it dies.

the formation clock
cosmic time
10⁻¹⁷ µs
horizon mass → PBH mass
10¹⁵ g
M_H ≈ 10¹⁵ g × (t / 10⁻²³ s) — collapse when a fluctuation enters the horizon
epoch
PBHs of ~10¹⁵ g

An overdense patch collapsing when the universe is 10⁻²³ s old traps roughly a horizon mass: ~10¹⁵ g. These are the holes whose explosions we could witness today.

Slide through cosmic time. Expansion continuously grows the horizon, so each epoch mints a characteristic PBH mass.

Nature would not mint a single mass but a distribution ψ(M), whose shape encodes the formation mechanism. The explosion rate we could hope to see today is set by how much of that distribution sits at the critical mass MU5.1×1014gM_U \simeq 5.1\times10^{14}\,\mathrm{g} — holes whose lifetime equals the age of the universe :

n˙PBH=ρDMψi(MU)3tU\dot{n}_{\mathrm{PBH}} = \rho_{\mathrm{DM}}\,\frac{\psi_i(M_U)}{3\,t_U}
mass function gallery
10⁹10¹¹10¹³10¹⁵10¹⁷10¹⁹10⁻²²10⁻²⁰10⁻¹⁸10⁻¹⁶10⁻¹⁴10⁻¹²PBH mass M [g]ψ(M) [normalized]M_U — exploding now
σ=0.1σ=0.5σ=1.0

Expected from a smooth, symmetric peak in the inflationary power spectrum. Width σ controls how much of the distribution leaks into constrained regions.

The explosion rate today is set by how much of ψ sits at M_U: ṅ = ρ_DM ψ(M_U) / 3t_U (Boluna et al. Eq. 3.8). For every allowed mass function it stays far below the HAWC bound of 3400 pc⁻³ yr⁻¹ — unless the distribution is spiked almost exactly at M_U.

Three physically motivated families, normalized and evaluated by the data pipeline (Boluna et al. Eqs. 3.9–3.14). The red line marks M_U: only the sliver of ψ crossing it explodes on our watch.
03dark degrees of freedom

A collider built by gravity

Here is the property that makes exploding black holes more than a curiosity: Hawking radiation is democratic. Gravity couples to everything, so a black hole radiates every particle species lighter than its temperature — quarks, neutrinos, gravitons, and anything else that exists, including particles that never touch our detectors .

The evaporation coefficient α(M) literally counts the particle content of nature. As the hole shrinks and heats, each new species unlocks like a threshold in a collider energy scan:

the degrees-of-freedom staircase
10⁶10⁸10¹⁰10¹²10¹⁴10¹⁶10¹⁸01234black hole mass M [g] (← hotter · lighter)α(M) — evaporation channels [rel.]e⁺e⁻μ⁺μ⁻u,d quarks + gluons (QCD)s quarkτ, cb quarkW,Zt, Higgs
Standard Model

Each step: the shrinking hole gets hot enough to radiate a new species (T = 1.058×10¹³ GeV·g / M). dM/dt = −α(M)/M², so more channels → faster evaporation. Dark degrees of freedom leave gravity no place to hide — they must be radiated too, whether or not they couple to light. Thesis §2.8.

α(M) rises each time k_BT crosses a particle mass (thresholds marked). Toggle the dark sector: doubling the degrees of freedom roughly halves the lifetime and dilutes the photon share of the luminosity — an observable, falsifiable shift.

In its final seconds a PBH reaches temperatures beyond any accelerator — a cosmic-scale collider whose luminosity, spectrum, and duration all depend on the full particle spectrum of nature. A dark sector with N copies of the Standard Model would shorten the final TeV burst by ~N and dim its photons by the same factor . Watching one explosion is a census of everything that can exist.

Even without new physics, evaporation carries a unique fingerprint. The measured power-law slope of the spectrum evolves as the hole shrinks, converging to a universal value that no astrophysical source reproduces:

the universal spectral index
10⁸10¹⁰10¹²10¹⁴10¹⁶0246black hole mass [g] (evaporation runs right → left)spectral index γ (dN/dE ∝ E^−γ)γ → 1.5 universal endpoint
pivot 1 GeVpivot 1 MeV

As the hole approaches its final moments, the measured power-law slope at any energy converges to γ = 1.5 — a fingerprint no astrophysical source shares. This is the primary template used to screen the Fermi transient catalog. (Boluna et al. Eq. 4.8, Fig. 6.)

Regenerated from the data pipeline (Boluna et al. Eq. 4.8, cf. Fig. 6). At 1 GeV the slope sweeps through γ ≈ 2–3 (quasar-like territory) before locking onto γ → 1.5 near death.
04the instrument fleet

Six eyes on the γ-ray sky

No single telescope covers the evaporation story. The spectrum spans nine decades of energy as the hole heats from MeV to TeV, so the search is inherently multi-mission : space telescopes (Fermi's GBM and LAT) own the keV–GeV band with huge fields of view; ground arrays (VERITAS, HAWC, LHAASO) own the TeV band with vast effective areas.

effective area atlas
10¹⁴ g
10⁻⁶10⁻⁴10⁻²110²10⁴10⁶110²10⁴10⁶10⁸10¹⁰photon energy [GeV]effective area [cm²]
GBMHAWCLATLHAASOVERITASE·dN/dE spectrum shape (arb. norm)
energy coverage
Fermi GBM
BATSE
Fermi LAT
VERITAS
HAWC
LHAASO
keVMeVGeVTeVPeV
Digitized instrument response curves from EffectiveAreas/*.dat in the BHRad repo (cf. Boluna et al. Fig. 3). Overlay the PBH spectrum and slide its mass: watch which instrument 'owns' each stage of evaporation.

Combining instruments honestly is its own craft. The analysis behind this site uses threeML , the multi-mission maximum-likelihood framework: every instrument keeps its own response and background model, while a single physical source model is fit jointly across all of them .

anatomy of a joint fit
FermiGBMBurstCatalog · TimeSeriesBuilder

Raw time-tagged photon events from each GBM scintillator and LAT tracker are pulled straight from the Fermi archives. Every detector has its own response, background, and quirks.

The threeML fitting pipeline used for every candidate in §06 — click through each stage.
05the brutal arithmetic of flux

Only the nearest explosions count

Here is the sobering part. A PBH explosion is intrinsically identical every time — same mass, same lightcurve, same luminosity. That makes the detection criterion pure geometry :

NS= ⁣ ⁣d2NdEdtAeff(E)4πd2dEdt    10orNSNB5N_S = \int\!\!\int \frac{d^2N}{dE\,dt}\,\frac{A_{\mathrm{eff}}(E)}{4\pi d^2}\,dE\,dt \;\geq\; 10 \quad\text{or}\quad \frac{N_S}{\sqrt{N_B}} \geq 5

Run that requirement through every instrument and you get a sensitivity frontier: the farthest distance each telescope could see a hole of a given mass.

the sensitivity frontier
10⁹10¹⁰10¹¹10¹²10¹³10¹⁴10¹⁵10¹⁶10⁻⁷10⁻⁵10⁻³10⁻¹10black hole mass [g]max visible distance [pc]0.1°1°10°
BATSEVERITASLATGBMHAWCLHAASOproper motion threshold

Anything below a curve is detectable by that instrument (N_S ≥ 10 photons or 5σ over background, 1-yr observation). Regenerated from the repo detectability pipeline (Analytical_Modelling.ipynb; cf. Boluna et al. Fig. 4). Above the dashed contours the source would visibly streak across the sky — a smoking-gun signature, but also a challenge for catalog association.

Regenerated end-to-end by this site's data pipeline from Analytical_Modelling.ipynb detectability code — same effective areas, backgrounds, and detection criteria (cf. Boluna et al. Fig. 4). Everything below a curve is visible to that instrument.

The peak reach is a fraction of a parsec — thousands of times closer than the nearest star. And nearness cuts both ways: anything that close should visibly move. At galactic speeds (~220 km/s), a source at 0.01 pc drifts about a degree per year — a smoking-gun streak, but also a reason catalogs might discard the very sources we want.

our cosmic backyard
Sun
Earth (1 AU)
Voyager 1 (~165 AU)
Oort cloud edge (~4000 AU)
Proxima Centauri
Barnard's Star
~30 nearest stars
1 AU10⁻⁴ pc10⁻² pc1 pc30 pc
search radius0.1 pc
volume · 4.2×10⁻³ pc³
expected PBHs near explosion (EGB-limited) · 1.7×10⁶

Detection horizons (◆) are the peak reach from the sensitivity frontier above. The tension in one picture: γ-ray telescopes only see explosions within a bubble much smaller than the distance to the nearest star, while abundance limits make such nearby events rare.

Log-scale map from Earth outward. The detection horizons sit deep inside the Oort cloud's neighborhood — set the search radius and see how many exploding holes the abundance limits allow inside it.
06template vs. reality

Fitting the flash

Suppose a short burst trips the GBM. Is it a neutron-star merger at a billion parsecs, or a black hole dying at a thousandth of one? The lightcurve holds the answer. The PBH template rises as (τt)0.52(\tau - t)^{-0.52} toward the explosion epoch τ — time-reversed compared to every conventional burst — optionally followed by an afterglow from the ejecta shell .

fit real bursts by hand
-200204001e+42e+43e+44e+4time since trigger [s]background-subtracted counts / 100 msτ (burst end)
GRB141222298 — real GBM data (100 ms bins)PBH template: (τ−t)^−0.52 + afterglow
burst epoch τ1.5s
log₁₀ norm3.20
afterglow spread12.0s
log₁₀ afterglow norm2.00

These are the actual light curves fitted in the thesis pipeline (Lightcurve_Fitting/Fitting.ipynb). In the real analysis, ultranest explores this exact parameter space against every GBM detector simultaneously via threeML — you are doing by hand what nested sampling does with 400 live points.

Background-subtracted 100 ms lightcurves for seven candidate GRBs, exported directly from Lightcurve_Fitting/~100ms_Source_Data in the BHRad repo. Drive the template parameters the Bayesian sampler explores.

In the production pipeline these fits run through threeML with ultranest's nested sampling — hundreds of live points exploring normalization, decay index, and afterglow timing against every GBM detector simultaneously, returning posterior distributions and Bayesian evidence for template comparison.

07mining fourteen years of Fermi data

The candidate hunt

The Fermi GBM has logged thousands of bursts since 2008. Almost all are ordinary. The search strategy applies three cuts inherited from the physics: the burst must be short (T90 within 0.2–5 s), hard (more high-energy fluence than low), and local — no measured redshift, since a real PBH burst comes from inside our stellar neighborhood.

The idea traces back to Cline's BATSE analyses in the 1990s , which found a curious subpopulation of very short, anomalously hard bursts. The repo's Fermi catalog search modernizes this on Fermi's much deeper catalog.

filter playground — real candidates
max T905.0s
min hardness1.00
passing cuts · 35 / 36
11011010²T90 [s]hardness (LAT fluence / GBM fluence)

Real sources from the repo's Fermi catalog search (H>1_T90[0.2-5]_RS=0.csv): each passed hardness > 1, T90 within 0.2–5 s in GBM or LAT-LLE, and no measured redshift. Sources marked ∞ had zero GBM-band fluence — the hardest events in the sample. Click a point to inspect it.

The 36 sources that survived the repo's full selection (H>1_T90[0.2-5]_RS=0.csv), with live re-filtering. Hardness here is LAT fluence over GBM fluence; Cline's original hardness–duration anticorrelation used BATSE bands.
08the isotropy test

Where do they point?

Geometry offers a free hypothesis test. Neutron stars, magnetars, X-ray binaries — everything born of stars traces the Milky Way's disk. But local PBH explosions sample only our tiny corner of the halo, so their sky map should be isotropic, indifferent to the galactic plane .

the candidate sky, in galactic coordinates
GCGRB160325291GRB150403913GRB141222298GRB160829334GRB120316008GRB120709883GRB101014175GRB141102536GRB150118409GRB090626189GRB180703949GRB120830297GRB090228204GRB160709826GRB090510016GRB110921912GRB131014215GRB110120666GRB140619475GRB080825593GRB190606080GRB140102887GRB101123952GRB110728056GRB140402007GRB120624933GRB171011810GRB150210935GRB191202867GRB220107793GRB081024891GRB190515190GRB120915000GRB130502327GRB170728961GRB160521385
north / south of plane · 37 / 39 (ratio 0.95)
toward / away from center · 28 / 48 (ratio 0.58)

Galactic coordinates, galactic center at the middle, plane along the equator. If these sources were neutron stars or other stellar remnants they would trace the disk; local PBH explosions should be isotropic. Sky positions from the candidate CSV and the 1FLT fitted-parameter table (cf. BoresightSelection.py, thesis Fig. 3.5).

Mollweide projection with the galactic center at the middle and the plane along the equator — the same visualization produced by BoresightSelection.py in the BHRad repo. Equatorial coordinates converted to galactic on the fly.

Cline additionally reported an anomalous cluster of very short bursts in one octant of the sky — never confirmed, never fully refuted. A larger candidate sample with Fermi-era statistics is exactly what could settle it.

09the multi-messenger encore

Echoes weeks after the flash

The γ-ray flash may not be the last word. The explosion dumps ~10²⁵ J of electron-positron pairs into the interstellar medium; Rees noted in 1977 that this conducting fireball plows the ambient magnetic field into a coherent low-frequency radio pulse — potentially detectable across the entire galaxy, far beyond the γ-ray horizon.

And blazar physics adds a slower channel: adiabatically expanding ejecta re-radiate at progressively lower frequencies, with radio peaking 40–140 days after the γ-ray flare in observed jets. Would a spherical PBH shell do the same? That is an open question the repo's afterglow notes flag explicitly .

one explosion, two messengers
γ-ray (ms)radio (weeks–years)
1 µs1 ms1 s1 hr1 mo30 yr
γFinal γ-ray flash

The last ~10⁹ g evaporate in about a microsecond–millisecond of TeV-scale emission. This is the 'burst' a GBM-like detector triggers on.

The observational strategy: catalog the position of every candidate burst, then watch those coordinates for a late-arriving radio transient. A match would be extraordinary — no astrophysical short GRB should produce this particular γ→radio sequence from a stationary point at parsec distance. Radio non-detections (ETA: < 2.3×10⁻⁷ pc⁻³ yr⁻¹, Cutchin 2015) already constrain the Rees–Blandford channel.

Click along the timeline. The γ-ray burst and the radio afterglow are separated by up to nine orders of magnitude in time — a coincidence search across archives is the discovery strategy.
10the slow-burn search

Sources that only brighten

There's a second way to catch a dying black hole: before the explosion. For months to years, a nearby PBH would appear as a faint GeV point source with a steadily rising flux :

Fγ(t)2.7×108(pcd)2(τts)0.533cm2s1F_\gamma(t) \simeq 2.7\times10^{-8}\left(\frac{\mathrm{pc}}{d}\right)^{2}\left(\frac{\tau - t}{\mathrm{s}}\right)^{-0.533}\,\mathrm{cm^{-2}\,s^{-1}}

The Fermi LAT Transient Catalog lists 35 sources with no known counterpart at any wavelength. The thesis transient-fitting pipeline fit every one of them with this two-parameter model .

fit a rising transient
10⁶10⁷10⁸10⁻⁷10⁻⁶time since first detection [s] (~decade of LAT data)flux > 0.1 GeV [cm⁻² s⁻¹]
mock unassociated transient (1FLT-like)your model: F ∝ d⁻² (τ−t)^−0.533
remaining lifetime τ12.6 yr
distance d2.82 mpc

Notice the degeneracy: raising τ and shrinking d can produce nearly identical decade-long light curves — exactly why the 33 fitted 1FLT sources carry wide error bars (TransientSources_fitted_params.csv). The tell that finally breaks the degeneracy is proper motion: at milliparsec distances the source should drift degrees per year, which the LAT catalog does not observe for these sources.

A mock decade-long LAT lightcurve with realistic scatter. Trade τ against d and feel the degeneracy that dominates the real fits — then check the frontier plot in §05 with the real fitted sources overlaid.

The verdict so far: the fits succeed — two clusters of solutions, one near 10¹² g at milliparsec distances, one near 10¹⁴⁻¹⁵ g further out — but both imply proper motion of degrees per year that the LAT catalog does not observe. The simplest reading is that these transients are something else. The method, though, now exists and sharpens with every year of data.

11scientific honesty

Where the evidence stands

No exploding black hole has been found. This site would be dishonest to imply otherwise — and dishonest to imply the search is hopeless. Here is the current ledger, claim by claim. Click to expand.

the evidence board

The deeper reason to keep looking: a single confirmed detection would simultaneously prove black holes evaporate (quantum gravity's only accessible prediction), demonstrate dark matter physics beyond WIMPs, and census every particle degree of freedom in nature — visible or dark. Few observations in physics carry that much payload.

The tools are improving on schedule: CTA will push the sensitivity frontier past every current instrument, LAT keeps accumulating exposure, and the template methods built in this research — universal lightcurves, spectral-index screening, transient fitting, proper-motion vetoes — are ready for the data.