Disentangling Free-Floating Planets from Stellar Flares in Roman Time-Domain Data Using Explainable AI
| Program ID | 19059 |
|---|---|
| Science Category | Exoplanets & Exoplanet Formation |
| Program Type | Analysis |
| Category | Small |
| Principal Investigator | Javier Viaña |
| PI Institution | Massachusetts Institute of Technology |
| Co-Investigators |
|
| Abstract | Free-floating planets (FFPs) are among the most elusive objects in the galaxy, but gravitational microlensing offers one way to find them. The Galactic Bulge Time Domain Survey of Roman is a unique opportunity to detect these worlds. However, a serious obstacle stands in the way: stellar flares can mimic FFP microlensing signal. Roman's color filters, which could in principle separate the two phenomena, will not be sampled with enough cadence to distinguish the shortest and scientifically most interesting events. We propose a pipeline to solve this problem. First, an EventFinder algorithm will scan Roman's F146 white-light time series for candidate events. Second, a machine learning classifier built on convolutional neural networks will distinguish flares from genuine FFP events. A secondary AI-based anomaly detection model will catch unusual events that fall outside the training data. The models will also incorporate explainable AI techniques so that the physical basis for each classification is available to the community. We will release public catalogs of FFP candidates, anomaly scores, explainability maps, the training data used, and open-source code developed. The result of this proposal will allow us to constrain the FFP mass function to an order of magnitude better than current surveys. |