The phenomenon of thermal diffusion (mass diffusion driven by a temperature gradient, known as the Ludwig–Soret effect1, 2) has been investigated for over 150 years, but an understanding of its underlying physical basis remains elusive. A significant hurdle in studying thermal diffusion has been the difficulty of characterizing it. Extensive experiments over the past century have established that the Soret coefficient, ST (a single parameter that describes the steady-state result of thermal diffusion), is highly sensitive to many factors3, 4, 5, 6, 7, 8, 9. This sensitivity makes it very difficult to obtain a robust characterization of thermal diffusion, even for a single material. Here we show that for thermal diffusion experiments that span a wide range in composition and temperature, the difference in ST between isotopes of diffusing elements that are network modifiers (iron, calcium and magnesium) is independent of the composition and temperature. On the basis of this finding, we propose an additive decomposition for the functional form of ST and argue that a theoretical approach based on local thermodynamic equilibrium3, 5, 10 holds promise for describing thermal diffusion in silicate melts and other complex solutions. Our results lead to a simple and robust framework for characterizing isotope fractionation by thermal diffusion in natural and synthetic systems.