What can we learn about initial disk conditions and planetesimal formation from ALMA disks? Here, we show our first steps toward a Bayesian retrieval of 15 parameters controlling a 1D viscous accretion disk model with embedded two-population dust. In order to achieve this using MCMC, we also have to introduce a surrogate model to speed the process up. This replacement of the physical model is a neural net with fully connected layers. We find (preliminary): Massive, compact initial disks with low viscous α, relatively weak external UV field and little dust drift to be favored by the framework.