Using neural networks to analyze and reconstruct images from gravitational microlensing events.Gravitational microlensing events produce images that are often noisy, blended with background light, and distorted by the telescope's point-spread function. Neural networks have become powerful tools for reconstructing the true signal from these messy observations. They can separate the lensing event from contaminating sources, estimate the lens parameters, and even identify anomalies that classical algorithms might miss.
The application of neural networks to microlensing analysis gained traction in the late 2010s, with studies demonstrating their ability to classify events and recover parameters faster than traditional chi-squared fitting.
A neural network is trained on large sets of simulated microlensing light curves and images, learning to map raw pixel data or photometric time series to the underlying physical parameters. Once trained, the network can process new observations in seconds, deconvolving the image, removing noise, and predicting the mass, distance, and relative motion of the lensing object. Some architectures also use unsupervised learning to detect unusual events without needing labeled examples.
Real-time analysis of microlensing alerts, such as those from the Roman Space Telescope's planned survey, will rely on neural networks to quickly identify and characterize events for follow-up observations.
Neural networks can reconstruct the image of a lensed source even when the event is so faint that the signal is buried in noise below the detection threshold of conventional algorithms.
| x | variable | data sample — sample |
| z | variable | latent variable sample — latent variable |
| D(x) | variable | discriminator output for real data — discriminator output |
| G(z) | variable | generator output for latent variable — generator output |
| V(D,G) | variable | value function of the game — value function |
| min_G | operator | minimize with respect to generator parameters — minimization over generator |
| max_D | operator | maximize with respect to discriminator parameters — maximization over discriminator |
| E_{x~p_data} | operator | expectation over the empirical data distribution — expectation over data distribution |
| E_{z~p_z} | operator | expectation over the latent prior distribution — expectation over latent distribution |
| log | operator | natural logarithm — logarithm |
