ML assisted magnetic properties optimization in classic soft magnetic alloy FeSiAl

V.A. Milyutin, R. Bures, M. Faberova
Institute of Materials Research, Slovak Academy of Sceinces,

Keywords: soft magnetic alloys, magnetic properteis, materials informatic


In our work, we develop a data-driven solution for functional properties optimization in soft magnetic alloy FeSiAl. Depending on the application, different properties of soft magnetic materials are put at the fore. The most important may be high saturation magnetization, low coercivity, high magnetic permeability, and, finally, high electrical resistivity, which is important for the minimization of eddy current losses at higher reversal frequencies of magnetization. The combination of Si and Al as alloying elements of iron opens up great opportunities for tailoring these properties over a wide range. Depending on the Si-Al ratio some properties may differ by an order of magnitude. This ternary system covers alloys with huge permeability, low coercivity, and high electrical resistivity. The development of FeSiAl alloys with required properties has so far only been carried out experimentally by trial and error based on basic rules of thumb. The problem with this approach is not only the increased time and expense required to achieve the desired result but also that many compounds remain unexplored, since at first glance they may not have interesting properties. To achieve this goal, we use supervised learning and solve regression problems. For this, we created a database using literature sources. It contains different compositions and results of experimental measurements of basic soft magnetic properties. Using this database we trained "classical" machine learning models such as random forest, gradient boosting, k-nearest neighbors, support vector machine, etc., and neural networks. We found that different models showed the best prediction ability for different properties (targets). Moreover, the applicability of the model depends on the range of compositions of interest. For example, despite the generally good values of basic metrics, tree-based models perform poorly in the area of high Al and Si content. This is due to the known problem that such models have limitations for extrapolation and predicting extreme values. Neural networks do not have this problem, however, due to the insufficient size of the database used for training, they suffer from a weak generalization ability. Therefore, to successfully solve the problem, we use a combination of several models. At this stage of work, we have already carried out a full cycle of work necessary for property prediction. Work is currently underway to improve the accuracy of predictions through the use of advanced models that have been actively developed in recent years. As a result, it becomes possible to choose a soft magnetic alloy with the desired properties necessary for a specific practical application.