Data Mining based development of Fe-Co-Ni soft magnetic material

Abstract

Accelerated materials development has gained interest in the research community as this approach can generate large amount of data in short period of time, and models built from this data can optimize the properties of the material for specific engineering applications. One such material is soft magnetic material which are widely used in electric motors, generators, transformers, inductors, etc. To model and optimize the properties, machine learning can be deployed. We present preliminary findings of the data mining of the literature for the Fe-Co-Ni alloy system. The processing conditions, properties, and compositions were obtained from the literature. The use of a machine learning tool to predict the properties based on composition, processing parameters, and structural information is described.Introduction: Globally, in the past two decades, total energy consumption has increased exponentially and electric motors account for a significant fraction of the total power consumption[1]. Soft magnetic materials play an important role in energy conversion and are widely used in electric motors, transformer cores, electric vehicles, inductors, etc [2]. However, commercially available soft magnetic materials do not have the optimum balance of properties needed in high frequency electric motors operating in extreme environments. Hence, machine learning based magnetic property prediction, optimization and design of soft magnetic materials has gained attention [3-5]. We build a Fe-Co-Ni based metallic soft magnetic materials database from literature and use a neural network algorithm [6, 7] to optimize properties.Database and Machine Learning: A database of 891 data entries comprising magnetic, mechanical, and electrical properties of various composition of Fe-Co-Ni alloys processed at various conditions was compiled from relevant literature. To curate this database, 55 references were compiled, including standard books, handbooks, and relevant publications. The atomic % of other elements in the alloys was restricted to maximum of 10%. Further, the database was preprocessed to convert the weight % to atomic % of the elements, to convert alphanumeric type of material and processing condition entries to numerical, and to calculate grain size from coercivity and yield strength of the alloys. This preprocessed database was used to train the machine learning tool with input variables: (a) composition, (b) type of material (categorical), (c) processing parameters like treatment temperature and time, cooling rate, % reduction in area (% r.a.) either by cold rolling or hot rolling, reheating temperature and time, etc., (d) grain size, and (e) volume fraction of phases (FCC, BCC, HCP). The output properties were saturation magnetization, log(coercivity), log(maximum permeability), log(electrical resistivity), yield strength, tensile strength, log(elongation), Vickers hardness, and cost of the material. The quality of fit, coefficient of determination (R2) values was determined through cross-validation. Using the trained model, properties of various compositions were predicted.Results and Discussion: Figure 1 shows the R2 values from cross-validation and the uncertainty in prediction for the output parameters. The algorithm was able to predict the properties in good agreement with the literature data. However, the prediction for electrical resistivity and elongation were not as good as that of other properties. This can be attributed to the lack of the data and complex behavior of the properties with the input parameters. The trained model was used to predict the properties of various Fe-Co-Ni alloys with the following conditions: cold rolled sheets with 95% r.a., annealed at 1000oC for 2 hours, and furnace cooled. The predicted saturation magnetization for the above specified conditions reveals a complex interplay between the properties, hence there it could be possible to design a soft magnetic material with improved mechanical strength.Conclusion: Using this robust neural network model and data mining, preliminary results relevant to the development of a Fe-Co-Ni based soft magnetic material with optimal combination of properties could be obtained.Acknowledgement: This work is supported by the AME Programmatic Fund by the Agency for Science, Technology and Research, Singapore under Grant No. A1898b0043, and the Royal Society. **