It is very common in astrophysics that certain relevant parameters of the objects studied cannot be obtained directly from observations, but require numerical models that simulate the relevant physical mechanisms, often through an iterative process of trial and error. This is often lengthy and consumes a large part of the human and material resources of the research process. This is the case in the characterisation of the circumstellar envelopes of Asymptotic Giant Branch (AGB) stars, which could greatly benefit from the use of Artificial Intelligence Deep Learning (DL) techniques. We present the preliminary results of this project as an illustrative example of what can be achieved by (and expected from) the application of DL to numerical modelling in astrophysics. For this project we have trained a convolutional neural network (CNN) to predict the physical conditions of these envelopes (mass loss, molecular abundance, temperature distribution, expansion velocity pattern, and distance) from a reduced set of single-dish observations of CO rotational transitions including $J$=1--0, $J$=2--1, as observed with IRAM 30m, and $J$=6--5, $J$=10--9, and $J$=16--15 as observed with HERSCHEL/HIFI. The network has been trained by building a complete library of numerical models covering a large range of parameters using a custom radiative transfer 1D code, which computes the excitation of CO using the Large Velocity Gradient (LVG) approximation and then solves the radiative transfer problem via ray tracing, thus generating synthetic CO profiles from a detailed description of the physical structure of a circumstellar envelope. Once the learning is completed, the system should be able to determine these fundamental parameters from observations without the need for 'manual' model fitting.%.