Gravitational microlensing is the only method sensitive to Earth-mass exoplanets and untethered free-floating worlds. Back in the 1980s Bohdan Paczyński showed that microlensing could hunt for dark objects, and today the Nancy Grace Roman Space Telescope is gearing up to survey 200 million stars toward the Galactic bulge. Roughly 1,400 bound exoplanets and 250 rogue planets are expected. To test formation and ejection scenarios, a complete sample in the shortest and faintest regime is needed — exactly where classical algorithms falter.
The architecture combines a shared transformer encoder that handles irregular time series via attention, with two specialized heads. The detection branch — an Evidence Network — trains on binary labels (signal/noise) using the l-POP exponential loss and outputs a calibrated Bayesian factor. The inference branch — Neural Posterior Estimation — employs a masked autoregressive flow to directly model the posterior. Both heads share the same latent representation extracted by the transformer from padded sequences (1,000 steps, three channels: time, flux, error). Training used 100 000 synthetic FSPL light curves generated via numerical simulation with augmentations: cadence gaps, point dropouts, and photometric noise. The whole pipeline is implemented with Machine Learning and simulation-based inference.
On a test set of 5 000 events, the Evidence Network at a threshold log₁₀K > 0.8 flagged 4 996 signals (99.9%) and zero false positives among 5 000 noise realizations (95% upper limit on FP < 6×10⁻⁴). By comparison, the hard cuts of Johnson et al. (2020), Sumi et al. (2011), and Mróz et al. (2017) yield 95.9%, 93.9%, and 95.4%, respectively — even when using oracle parameters to compute the statistics. The largest advantage emerges at ρ ≳ 5, where the fraction of recovered free-floating planet events is ~95% versus ~65% for the hard cuts. The posterior parameter distributions are well calibrated (TARP test), and the whole procedure takes ~21 ms per event — a ~16,000× speedup relative to MCMC.
Replacing deterministic thresholds with probabilistic Bayesian factors lets one flexibly tune the balance between completeness and purity to match the science goal. Crucially, the method nearly eliminates the blind spot for strong finite-source events — the very regime where the population of free-floating planets with masses down to Earth’s builds up. Calibrated posterior estimates for each event open the door to hierarchical analysis of the Galactic object distribution.
The method can be extended to binary lenses, which generate more intricate caustic patterns. Another direction is to incorporate astrophysical false positives (flares, eclipsing variables) and realistic noise into training for actual Roman data. A sliding window for long-duration events and hierarchical Bayesian population inference are also on the roadmap.
The technique applies to the entire class of transient searches on irregular time series — from supernovae to variable stars — where rapid, uncertainty-controlled decisions are needed.
Immediate next steps include validation on real data from existing surveys (OGLE, MOA) and adaptation to Roman data engineering.
The work addresses the unsolved problem of the formation and prevalence of free-floating planets, and more broadly the challenge of automated transient classification in large-scale surveys.
This section is addressed to the author of the paper and their colleagues. Neighbouring works were found by meaning-based search across our archive, not by word overlap. We do not review: we show the field around the work.
What the work does
The authors built a neural network pipeline for detecting and estimating parameters of gravitational microlensing events. A transformer encoder processes irregular time series, while two heads—an Evidence Network and Neural Posterior Estimation—output a Bayes factor and posterior parameter distributions. On synthetic data, the method achieves near-perfect recall, especially for events with strong finite-source effects, and speeds up parameter estimation by orders of magnitude compared to MCMC.
Where its strength is
Combining detection and inference in a single architecture with a shared representation allows flexible tuning of the precision–recall balance, while calibrated posterior distributions and Bayes factors provide direct probabilistic error control. This is especially valuable for short and faint events, where classical threshold-based methods lose sensitivity.
Where it could go next
Worth trying
Why this may matter
If the proposed directions maintain high recall and precision on real data, this could bring us closer to a reliable census of the population of free-floating Earth-mass planets and to testing scenarios of their formation and ejection.
Nearby works in our archive
The closest work in the archive sits at 0.64 by meaning — the usual level for our corpus.
The shortest microlensing event from a free-floating planet may last only a few hours, yet our sentinel catches it.
"Rogue Planet" by Greg Bear: wandering worlds that could have been found with the technique from this paper.