A pair of microcalorimetric sensors (pellistors), one with a hexamethyldisiloxane (HMDS) pretreatment, is applied for the determination of the composition of butane/methane mixtures in air. Both gases show a strong interaction on the sensor surface, excluding any linear superposition algorithms. An evaluation of the butane and methane concentrations from the pellistor signals by two neural networks, one for each combustible, is described here. Maximum errors below 4% LEL (lower explosion limit) for butane and 10% LEL for methane are attained with the neural networks being trained with 21 measured points and applied to 64 points.