The study of the composition of astrophysical ices is relevant because depending on the compounds present in the molecular clouds, it is possible to predict its future evolution. Due to the difficulty of obtaining real satellite data, the process is simulated in a laboratory synthesizing ice analogs formed by different combinations of molecules. The data correspond to the infrared spectra of ices. The spectra and abundances are non-negative, allowing a non-negative matrix factorization (NMF) of them. Because of the spectra of ices have supergaussian statistics, it is possible to apply a sparseness restriction. We review some NMF algorithms based on different divergence measures. Moreover, we analyze the modified versions of the algorithms imposing sparseness restriction. We perform several simulations in order to measure the improvement due to the imposition of the sparseness restriction. Besides, we check the performance of the algorithms under noise conditions. We present the results carried out with an ice analogs database, confirming the suitability of the NMF approach.