The Hunter of Invisible Worlds ✛ ✎

Original: "Microlensing Detection and Inference via Learned Bayes Factors"
arXiv:2607.20260v1 · 2026-07-22 · arXiv non-exclusive · independent analysis · 3 min
A neural network pinpoints rogue planets in a split second from the brief brightening of a background star.

Two hundred million stars. At the very center of our Galaxy, the future Nancy Grace Roman telescope will keep watch for fleeting flashes. These flashes occur when an invisible rogue planet — a free-floating planet (a world adrift by itself) or an exoplanet (a planet orbiting a distant star) — briefly amplifies the light of a background star with its gravity. The effect is called microlensing (gravitational focusing of light). Such events last only hours and are easy to miss.

Old search methods worked like a sentinel with binoculars: if a flash exceeded a certain threshold in a few photometric (brightness-related) features, it was flagged as a detection. But short flashes from low-mass planets often fell below the threshold and went unnoticed — up to 40% of events were lost. The new approach gives the sentinel a map. This is a neural network (a trained program) that doesn't just give a 'yes/no' answer but evaluates the full picture: how likely the flash is real, and what are the properties of the planet that caused it.

The shortest flash from a rogue planet can last just a few hours — quicker than the break between a sentinel's shifts.

Here's how the map works. Its core is a transformer — an architecture that processes uneven sequences of measurements (time, brightness, error) and uses an attention mechanism to highlight what is important. On this foundation, two 'heads' operate. The first learns to distinguish signal from noise using examples and outputs a Bayesian (probabilistic) factor — a number indicating how many times more likely it is that we are seeing a real lens. The second head uses the same landscape to instantly estimate the planet's parameters: mass, distance, speed. All training is based on Machine Learning (the computer learns from simulated data) and numerical simulation (computer calculations) of hundreds of thousands of artificial light curves.

Evaluating one event takes 21 milliseconds. On the same processor, the classical method would slog through an hour and a half — a speedup of 16 000 times.

Tests on 5 000 events showed that the map catches 99.9% of signals, with practically no false positives (fewer than one in two thousand). The method is especially good with the shortest and faintest events, which previously were simply lost. Now every captured wanderer immediately gets a 'passport' with no further calculations needed, allowing us to build an accurate census of the Galaxy's dark population. The sentinel with the map awaits the stream of data from the Roman telescope. And then, in the center of the Milky Way, where once we saw only flickering noise, it will trace sharp lines of invisible worlds.

The shortest flash from a rogue planet lasts just a few hours — less than a typical night's sleep.

In Greg Bear's novel 'Rogue Planet,' wandering worlds pose a threat; the new method turns them from a frightening unknown into subjects of precise science.

K = \frac{p(x \mid \mathcal{M}_1)}{p(x \mid \mathcal{M}_0)}
The likelihood ratio of the signal and noise models after accounting for all possible parameter values.
Scientists
Isaac NewtonJohn MichellRobert HookeDavid CharbonneauBohdan Paczyński
Concepts
bayesian evidence neural network microlensing reconstruction gravitational microlensing bayes factor Bayesian inference nested sampling maximum likelihood maximum likelihood estimation predictive learning neural networks event selection metropolis hastings algorithm machine learning materials pattern recognition gravitational lensing modeling astrophysical signal detection dark matter lensing anomaly detection prediction error posterior distribution importance sampling