Data Mining Approach to Development of Iron-Silicon Soft Magnetic Alloys

Abstract

There has been renewed interest in recent years to accelerate materials development beyond traditional approaches. A combination of both machine learning and high-throughput experiments can be used for predicting material properties and designing compositions beyond what is currently known. Herein we report the use of a computational tool, Alchemite™, to mine for important processing parameters of Fe-Si-Al alloys from a data set curated from published experimental results. The preliminary calculations suggest that this is a viable method to develop new Fe-Si based alloys.Introduction: Silicon steels (Fe-Si) are an important class of materials as they are widely used in applications like electric motors and transformers. Currently, 3.2 wt.% Si is widely used due to (i) good balance of electrical and magnetic properties, and (ii) ductility requirements in traditional processing methods.1 Recent trends have pointed to higher Si content of up to 10 wt.%, with an addition of small amounts of aluminum (Al) to improve material performance. Due to the small literature, it is imperative that we use machine learning to explore higher Si compositions and their processing methods to obtain a good balance of soft magnetic properties, electrical properties and mechanical properties.Data Set and Machine Learning Model: A typical machine learning model comprises three parts: training data, attributes for describing the material, and the machine learning algorithm to map the attributes to properties. The Fe-Si-Al data set was compiled of reported experiment values from 40 references (IEEE Ferromagnetism Handbook2 and 39 journal references, published between 2000 – 2020), resulting in a total of 572 entries. The attributes that were selected as input parameters to describe the material were (i) Composition, (ii) Type of material – like ingots, sheets or powder etc., (iii) Final thickness of the material, heat treatment conditions (iv) Treatment temperature, (v) Treatment time and (vi) Cooling rate, (vii) Frequency and (viii) Magnetic flux density. The output parameters comprised various soft magnetic, electrical and mechanical properties.Alchemite™ is a machine learning technology commercialized for materials and industrial chemicals design through Intellegens.3 The artificial neural network trains from the data set provided and predicts the selected material properties. The coefficient of determination (R2) values was obtained to determine the consistency of the data and predictions.Results and Discussion: Figure 1 shows the R2 values for the input parameters. The machine learning tool showed good predictions for most properties, although mechanical properties were not as well predicted (R2<0.8). This is expected as there is a gap in the reported literature. More work will be done to address this gap, by means of improving the data set through theoretical calculations and high-throughput experimental methods.We studied the machine learning model parameters to reveal the most important variables for each output parameter: The frequency of operation affects total core loss greatly, like observations in alternating current (AC) applications – core loss is higher at increased operating frequency. To mitigate such effects, electrical resistivity must be increased – by controlling material thickness, grain size, and composition.Next, the type of material strongly affects the coercivity. The coercivity is affected by the grain size and composition, and this is in turn is affected by the method of fabrication and subsequent processing.The tensile strength is sensitive to the cooling rate because it affects the formation of different phases (A2, B2, and D03). It is reported that the ordering of phases (in B2 and D03) can interact with dislocations, resulting in a strengthening effect which decreases ductility.1 The phases formed also depend on the Si content, as reflected by the correlation between yield strength and Si content.The above highlights how the various input attributes are interrelated in affecting the output properties, and physical insights this can deliver. This underpins the main difficulty in traditional trial and error experiments and the mammoth task to undertake if it is to be done manually.Conclusions: The results show that the use of Alchemite™ for the development of Fe-Si-Al alloys are robust and promising, as it can capture property-processing relationships and highlight gaps in the data set. More work will be focused on improving the data set and high-throughput experiments performed to compare the results obtained with the predicted values. This will enable us to add to knowledge on achieving optimized and balanced properties in Fe-Si based alloys.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. **