Digital Polisher: How a Neural Network Unveils the Secrets of Binary Gravitational Lenses ✛ ✎ ✦
Gravitational microlensing is a fascinating cosmic effect predicted by Einstein and developed by Zwicky. A massive body passing between us and a distant star bends its light like a lens, creating characteristic magnifications and distortions. When the lens is binary, a star with an exoplanet or a close pair of stellar remnants, complex caustic patterns flash on the photometric light curve, as if light were refracted through faceted crystal. Decoding these patterns means recovering masses and orbits, but until recently, each such analysis required hours of computation and thousands of simulations.
CausticFlow reimagines this task as the art of optical polishing. Instead of blindly searching parameters, the neural network learns to "feel" the lens geometry from the subtlest bends in the light curve. At its core are neural controlled differential equations (Neural CDE). They read the time series as a continuous stream, unaffected by gaps and irregularities, just as a skilled musician catches the melody even when notes drop out. Then, machine learning with normalizing flows transforms a simple distribution into a complex posterior, where each peak corresponds to a possible lens configuration. Training is done on numerically simulated light curves with noise matching real telescopes, and all computations are accelerated by the JAX library.
On synthetic tests, the accuracy is impressive: typical error in mass ratio q is about 17%, and in projected separation s only 3%. However, the model does not stop at a rough estimate: subsequent fast Nelder–Mead optimization with a dozen starting points drawn from the posterior polishes the parameters to better than 5% and 1% accuracy, respectively. On 10 real KMTNet events, despite parallax and orbital motion, CausticFlow successfully recovers 7 objects, spending no more than 10 minutes of CPU time per event. Where MCMC sampling would get bogged down in degenerate solutions, the neural network confidently finds all modes.
CausticFlow does not replace detailed Bayesian analysis but acts as an optical sight: it instantly points to promising regions in parameter space, turning hours of computation into seconds. For future missions like JWST and Hubble, this opens the possibility of processing millions of microlensing events in real time. The neural network scans the data stream from telescopes like a grandmaster analyzing an opening, selecting candidates for immediate spectroscopic follow-up in a split second. This shifts the hunt for dark matter and cold planets into a mass-production mode.
This work proves for the first time that neural network emulators can handle degenerate inverse problems in astrophysics. In the future, CausticFlow will learn to account for parallax and orbital parameters, and its architecture will become a template for analyzing irregular time series in other sciences. Just as good optics lets us see distant galaxies, this digital polisher will reveal a previously invisible universe — thousands of orphan planets and massive compact objects, and even rewrite the history of star formation in our Galaxy. And perhaps, among the millions of selected light curves, an anomaly will one day flash that forces us to revise the very equations of gravity.
Neural CDE reads the light curve as a continuous stream — much like we understand text with missing letters, recovering meaning from context. Thanks to this, CausticFlow successfully processes events where a ground-based telescope only captured part of the flare.
The technology echoes the idea from "The Three-Body Problem"—hunting hidden objects by minute distortions of light. CausticFlow does the same, but instead of alien civilizations, it searches for dim stellar remnants and rogue planets.