This paper presents a comparison between direct measurements and inverse predictions of a molten salt bank formation inside an experimental setup representing a real metallurgical reactor. This new inverse method rests on a quasi-2D virtual approach based on the adjoint problem and on the conjugate gradient method. A specific small-scale experimental setup was designed to support experiments conducted with high temperature and corrosive molten salts. Non-intrusive temperature measurements were taken, by means of an infrared camera, to feed the inverse procedure. Direct measurements of the experimental solidification fronts were done to validate the inverse method. Usual difficulties faced in the implementation of the inverse method in a real case are highlighted. It was found that, under the experimental setup conditions, the inverse heat transfer method leads to accurate predictions of the time-varying shape of the molten salt bank, i.e., accuracy is below 5 mm (domain size: [0-35 mm]). Results indicate however that a direct measurement of the heat flux released by the crucible on the bottom boundary is imperative to guarantee the effectiveness of the inverse method. Crown Copyright (C) 2012 Published by Elsevier Ltd. All rights reserved.