Hybrid Neural Network Modeling of Mechanical Alloying Process Parameters and Alloy Performance Using Forward and Reverse Mapping

ABSTRACT Mechanical alloying (MA) process synthesizes a nanostructured Ti‐Mg‐Zr ternary alloy suitable for biomedical applications. The desirable characteristics of titanium, magnesium, and zirconium (such as high biocompatibility, corrosion resistance and low density) enable the development of alloys with minimal magnetic resonance imaging (MRI) artifacts. Retaining the nanostructure (alloy performances such as crystallite size CS and lattice strain LS) in Ti‐Mg‐Zr alloys is influenced by MA parameters (milling speed: MS, duration: MD, and ball‐to‐powder‐weight ratio: BPR). The present work establishes forward and reverse modeling between the MA parameters and alloy performance using neural networks such as artificial bee colony (ABC‐NN), backpropagation (BPNN), and genetic algorithm (GA‐NN). Forward modeling predicts alloy performances from known MA parameters, while reverse modeling estimates the MA parameters for the desired alloy performances. A dataset of 1,000 samples was generated using empirical regression equations and was supplied to networks in batch mode. A detailed parametric study was conducted to optimize network parameters. The optimized network parameters resulted in reduced mean squared error (MSE) corresponding to forward and reverse modeling tasks equal to 0.0021 and 0.0065 for ABC‐NN, 0.0056 and 0.0097 for BPNN, and 0.0034 and 0.0076 for GA‐NN, respectively. The trained networks are tested against 10 randomly generated experimental cases for forward and reverse predictions, resulting in 3.41% and 3.55% for BPNN, 3.24% and 3.30% for ABC‐NN, and 3.30% and 3.33% for GA‐NN, respectively. The ABC‐NN predictions help industry personnel to scale up the synthesis of nanostructured Ti‐Mg‐Zr alloy for permanent implants requiring regular MRI examinations.