Introduction
Gravitational microlensing occurs when a foreground "lens" star passes in front of a distant "source" star, bending its light and causing a temporary brightening. When the lens is a binary system—such as a star and an orbiting planet—the resulting light curve becomes significantly more complex. Analyzing these two-lens, one-source (2L1S) events is a major challenge in modern astronomy.
Currently, researchers face two primary hurdles: the high computational cost of evaluating mathematical models for light curves and a pathological parameter space. The "likelihood surface" often contains narrow, deep local minima that cause standard sampling methods like Markov Chain Monte Carlo (MCMC) to fail. Consequently, experts must manually perform grid searches to find viable starting points. With the upcoming Roman Space Telescope expected to discover thousands of such events, this manual approach is no longer scalable.
Methods: Amortized Neural Posterior Estimation
To address these challenges, the authors propose a framework based on Amortized Neural Posterior Estimation (ANPE). By performing the heavy computational work during an upfront training phase, the system enables nearly instantaneous inference on new data.
The architecture consists of two primary components:
- The Featurizer: An 18-layer 1D Residual Network (ResNet) combined with a 2-layer Gated Recurrent Unit (GRU). This network compresses high-dimensional raw light curve data into a 256-dimensional feature vector.
- The Neural Density Estimator (NDE): A Masked Autoregressive Flow (MAF) with 20 blocks. This normalizing flow learns to transform a simple base distribution into the complex, multi-modal posterior distribution of physical parameters, such as mass ratio ($q$) and binary separation ($s$).
Datasets and Training
The model was trained on a massive synthetic dataset generated using the MulensModel software. The researchers simulated 1 million 2L1S events tailored to the expected performance of the Roman Space Telescope (15-minute cadence over 72-day windows). After filtering out events that were indistinguishable from simple single-lens events, a final dataset of 291,012 simulations was used, with 95% allocated for training and 5% for testing.
Key Insights and Results
A defining feature of this approach is its ability to handle mathematical "degeneracies." In microlensing, different physical configurations—such as "wide" versus "close" binary separations—can produce nearly identical light curves. The NDE successfully captures these ambiguities by producing multi-modal posteriors, representing multiple distinct peaks in probability.
The system achieved remarkable performance:
- Inference Speed: The model generates $10^5$ posterior samples per second on a single GPU, orders of magnitude faster than traditional methods.
- Accuracy: Parameters like binary separation ($s$) and mass ratio ($q$) were well-constrained in nearly 100% of test cases.
- Automation: The resulting posteriors are accurate enough to serve as a "warm start" for traditional MCMC samplers, effectively automating the pipeline and removing the need for manual human intervention.
Limitations and Future Work
While highly effective, the model has limitations. The parameter $\rho$ (source star size) was constrained in only 14% of cases, as its effect is only visible when a source passes extremely close to a caustic. Additionally, the current noise model is simplified (Gaussian/Poisson). Future iterations will incorporate more realistic, image-based simulations that include detector artifacts and crowded star fields to better prepare for real-world Roman Space Telescope data.
Summary updated