Inverse Neural Network Approach for Optimizing Chemical Composition in Shielded Metal Arc Weld Metals.

This study presents a hybrid machine learning framework combining an artificial neural network and a genetic algorithm to optimize chemical compositions of shielded metal arc weld metals for achieving targeted mechanical properties. First, a neural network model was trained using a large experimental database provided by Dr. Glyn M. Evans, which includes the chemical compositions and mechanical properties of over 950 shielded metal arc weld metals. The neural network model, optimized via Bayesian optimization, demonstrated high predictive accuracy for properties such as yield strength, ultimate tensile strength, and Charpy impact transition temperatures. To enable inverse design, a genetic algorithm-based optimization was applied to the trained neural network model, iteratively exploring the composition space to find optimal elemental combinations that match predefined mechanical property targets. The proposed hybrid approach successfully identified multiple feasible compositions that closely match the desired mechanical behavior, demonstrating the potential of neural network-assisted inverse design in welding alloy development.