A Fourier neural operator speeds up the modelling of protoplanetary discs by a billion times

An emulator trained on RADMC-3D spectra enables Bayesian inference on edge-on discs and the study of their ices at the population scale.

The protoplanetary disks observed by the edge show how the ices that will serve as the building blocks of the planets are distributed vertically. But deducing the abundances from these requires calculations of radiative transfer with multiple anisotropic scattering, which cost from 10 to 1,000 CPU hours per model: enough to preclude any rigorous Bayesian inference and any homogeneous study on many objects.

A median error of less than 5%

A team trained a Fourier neural operator on a grid of spectra calculated with RADMC-3D. The network reproduces these spectra with a median error of 4.84% while accelerating their evaluation by a factor of 10⁹ to 10¹⁰. Integrated into a nested sampler, it finds the parameters of synthetic disks without bias.

Applied to the ESO-Hα 569 disk, it concludes that the disk is massive, almost edge-on (about 80°), in agreement with previous work, even if the abundances of ice remain unconstrained by the spectral distribution of energy alone. The authors see it as a tool to exploit the large JWST surveys, while recommending that the key results be verified with complete models.