Self-organizing maps for pattern recognition in design of alloys

Author(s): 
Rajesh Jha, George S. Dulikravich, Nirupam Chakraborti, Min Fan, Justin Schwartz, Carl C. Koch, Marcelo J. Colaco, Carlo Poloni & Igor N. Egorov (University of the Witwatersrand Johannesburg, South Africa)

CHALLENGE - AlNiCo are permanent magnetic alloys that have been widely used due to affordability, high-temperature stability, and excellent anticorrosion properties. A novel approach is presented for creating a work plan using a set of computational tools for the design of alloy chemistry and multi-objective optimization of desired macroscopic properties of various alloys.

SOLUTION - Chemical concentrations of eight alloying elements were initially generated using a random number generator so as to achieve a uniform distribution in the design variable space. Meta-models that were selected were then used to simultaneously maximize the properties while minimizing cost of the raw materials and alloy mass density. It was decided to Pareto-optimize chemical compositions of alloys for the three optimized properties ((BH)max, Br, and Hc) and then use these compositions to predict the other seven properties. Unsupervised learning algorithms such as principal component analysis (PCA), hierarchical clustering analysis (HCA), and self-organizing maps (SOM) were used to discover various patterns within the dataset. In this work, modeFRONTIER is used for SOM analysis.

BENEFITS - Results show that alloy 124 predicted by this design methodology is the best alloy according to the three optimized properties. The presented approach can prove to be useful for designing new alloys and aiming for their accelerated implementation.