The Still Stone Amid Dancing Mirages: How to Find an Exoplanet Among Illusions ✛
A pebble on a streambed lies still as sunlight sparkles dance on the water's surface. In astronomy, the contrast is similar: a real exoplanet is the stone, while myriad speckles (patches of scattered starlight) are the ripples trembling in instruments. The most advanced coronagraphs diligently block the blinding stellar disk, but residual light spawns deceptive mirages that have fooled observers for decades. A planet and a speckle look nearly identical in a single image. To tell the still from the shifting, a different approach is needed.
Coherent Differential Imaging (CDI) turns speckles into a dance. Using a deformable mirror, a flexible membrane with hundreds of actuators, known distortions are introduced into the telescope's wavefront. Speckles, being coherent (their light can interfere, like a laser), react sensitively to each nudge, while the planet, whose light is incoherent, remains unmoved. The method requires a series of frames: on each, the mirror is slightly deformed differently, and the speckle pattern comes alive. This resembles Pair Wise Probing: sending pairs of opposite distortions, push and release, and from the responses discerning clear water from mere reflections.
The new algorithm EPICX, developed at the LIRA laboratory of Paris Observatory, employs gradient boosting (the XGBoost ensemble of decision trees). For three different coronagraphs — an ideal one, a phase mask, and the hybrid Lyot coronagraph of the future Roman telescope, named after Nancy Grace Roman—2,250 simulated scenes were generated using the Asterix package. Into each scene, 0 to 2 planets were injected with contrasts ranging from 2×10⁻⁶ to 10⁻³. Then all suspect points were run through the DAOStarFinder algorithm, and for each, a 12-layer feature cube was built capturing how brightness changes from push to push. XGBoost was trained on 1,200 scenes to classify candidates, prioritizing accuracy — since redirecting telescopes to check false alarms is prohibitively expensive.
On the remaining 750 scenes, EPICX separated wheat from chaff almost flawlessly. At a dynamic threshold maximizing the F0.5 metric, out of thousands of speckles only a handful were mistaken for planets, while about 80% of real worlds were detected — even those completely buried in the original image. Remarkably, the method worked equally well with both weak (λ/16, about 36 nm) and strong (λ/3, nearly 200 nm) mirror pushes, exceeding the linear CDI regime. The probability map from XGBoost highlighted the planet as a sharp peak, discarding the dancing halo of glints. A short series of 5–7 frames proves enough to spot the still stone.
The future looks exciting. EPICX will allow telescopes like Roman and the ground-based SPHERE+ to spend precious time not on speckle calibration but on studying the planets themselves and their atmospheres. In effect, the method paves the way from simple detection to spectroscopy of distant worlds — a task envisioned by Michel Mayor, the discoverer of the first exoplanet around a Sun-like star. Ahead lie more complex simulations with photon noise and tests on real skies. Perhaps one day we will see not just a solitary stone, but a whole necklace of planets frozen in a dance around their parent star.
The word ‘speckle’ comes from an Old English root meaning ‘little spot.’ These devious little spots have long taunted astronomers by posing as planets at every opportunity. To outsmart them, EPICX borrows an idea from another era: Pair Wise Probing was first proposed in 2007, long before anyone dreamed of widespread machine learning in astronomy.
In Carl Sagan’s novel *Contact*, Ellie Arroway struggles to fish a signal out of the noise — much like EPICX pulls a planet from a chorus of stellar interference, only with light waves instead of radio waves, and giant ground- and space-based observatories instead of radio telescopes.