2025 workshop program
Inferring Planet and Disk Parameters from Protoplanetary Disk Images Using a Variational Autoencoder
The work
Abstract
We introduce the first application of Variational Autoencoders (VAEs) for the physical interpretation of protoplanetary disk images, offering a new generative approach for inferring disk and planetary properties directly from ALMA dust-continuum observations. Our model is trained on a large suite of high-resolution FARGO3D + RADMC-3D simulations, capturing a diverse range of disk morphologies shaped by embedded planets. The VAE learns a physically disentangled latent space capable of reconstructing images with >99.5% SSIM, while jointly predicting the masses and orbital radii of up to three planets, along with disk parameters such as α-viscosity, Stokes number, dust-to-gas ratio, and flaring index—all with R² > 0.9. Applied to 23 real protoplanetary disk systems, the model recovers parameter estimates consistent with existing literature while uncovering new correlations across disk properties via the learned latent structure. This work establishes VAEs as a powerful tool for probabilistic, interpretable, and scalable parameter inference in protoplanetary disk science, bridging simulation and observation in a fundamentally new way.