Machine learning-based mass density model for hard magnetic 14:2:1 phases using chemical composition-based features

Abstract

The Fe14Nd2B-based permanent magnets are technologically sought-after for energy conversion due to their unparalleled high energy product (520 kJ/m3). The basis for the excellent magnetic properties of the material is the 14:2:1 intermetallic phase with its outstanding intrinsic properties. Depending on the desired property portfolio, different chemical compositions of 14:2:1 phases are used. With such 14:2:1 phases, the conversion of literature data and/or measured magnetic moments (in µB/f.u. or emu) into the magnetic saturation polarization (in Tesla) is often a challenge because the mass density, required for this, frequently does not get reported. We present a ‘machine learning’ mass density model for 14:2:1 phases, using chemical composition-based features (representing 33 elements). The datasets for training and testing contain 189 phases (176 compositionally different) with their literature reported densities. The model with compositional features achieved a low mean-absolute-error of 0.51 % on unseen test-dataset.