Abstract Five hundred eighty-five viscosity measurements on 40 melt compositions from the ternary system CaMgSi 2O 6 (Di)-CaAl 2Si 2O 8 (An)-NaAlSi 3O 8 (Ab) have been compiled to create an experimental database spanning a wide range of temperatures (660–2175°C). The melts within this ternary system show near-Arrhenian to strongly non-Arrhenian properties, and in this regard are comparable to natural melts. The database is used to produce a chemical model for the compositional and temperature dependence of melt viscosity in the Di-An-Ab system. We use the Vogel-Fulcher-Tammann equation (VFT: log η = A + B/(T − C)) to account for the temperature dependence of melt viscosity. We also assume that all silicate melts converge to a common viscosity at high temperature. Thus, A is independent of composition, and all compositional dependence resides in the parameters B and C. The best estimate for A is −5.06, which implies a high-temperature limit to viscosity of 10 -5.06 Pa s. The compositional dependence of B and C is expressed by 12 coefficients (b i=1,2.6, c j=1,2..6) representing linear (e.g., b i=1:3) and higher order, nonlinear (e.g., b i=4:6) contributions. Our results suggest a near-linear compositional dependence for B (<10% nonlinear) and C (<7% nonlinear). We use the model to predict model VFT functions and to demonstrate the systematic variations in viscosity due to changes in melt composition. Despite the near linear compositional dependence of B and C, the model reproduces the pronounced nonlinearities shown by the original data, including the crossing of VFT functions for different melt compositions. We also calculate values of T g for melts across the Di-An-Ab ternary system and show that intermediate melt compositions have T g values that are depressed by up to 100°C relative to the end-members Di-An-Ab. Our non-Arrhenian viscosity model accurately reproduces the original database, allows for continuous variations in rheological properties, and has a demonstrated capacity for extrapolation beyond the original data.