Aluminosilicate glasses and melts are of paramount importance for geo-and materials sciences. They include most mag-mas, and are used to produce a wide variety of everyday materials, from windows to smartphone displays. Despite this impor-tance, no general model exists with which to predict the atomic structure, thermodynamic and viscous properties of aluminosilicate melts. To address this, we introduce a deep learning framework, 'i-Melt', which combines a deep artificial neu-ral network with thermodynamic equations. It is trained to predict 18 different latent and observed properties of melts and glasses in the K2O-Na2O-Al2O3-SiO2 system, including configurational entropy, viscosity, optical refractive index, density, and Raman signals. Viscosity can be predicted in the 10(0)-10(15) log(10) Pa.s range using five different theoretical frameworks (Adam-Gibbs, Free Volume, MYEGA, VFT, Avramov-Milchev), with a precision equal to, or better than, 0.4 log(10) Pa.s on unseen data. Density and optical refractive index (through the Sellmeier equation) can be predicted with errors equal or lower than 0.02 and 0.006, respectively. Raman spectra for K2O-Na2O-Al2O3-SiO2 glasses are also predicted, with a rel-atively high mean error of similar to 25% due to the limited data set available for training. Latent variables can also be predicted with good precisions. For example, the glass transition temperature, T-g, can be predicted to within 19 K, while the melt configu-rational entropy at the glass transition, S-conf(T-g), can be predicted to within 0.8 J mol(-1) K-1.