Deep learning is considered as a disruptive method in the field of mineralogy and hyperspectral imaging. Many techniques exist to gain mineralogical information. Amongst them powder X-Ray diffraction (XRD) is very popular and powerful, while hyperspectral imaging is used in many applications such as Earth observation. A key issue for both XRD and hyperspectral imaging is not only to identify the endmembers constituting a mixture but also quantify the abundance of each endmember. In this study, we propose completly novel neural network (NN) training losses specifically designed for proportion inference. Extensive experiments illustrate that the proposed approach allows validated NN architectures to be trained to infer accurately on proportions.