Abstract
機械構造用材料の力学的性質の測定精度および効率を向上させるために、3層パーセプトロン型のニューラルネットを用いて、機械構造用Cr-Mo鋼、Ni鋼、Ni-Cr鋼およびNi-Cr-Mo鋼の降伏点、引っ張り強さ、伸び、絞り、衝撃値、硬さについての推定を行った。機械構造用Cr-Mo鋼、Ni鋼、Ni-Cr鋼およびNi-Cr-Mo鋼の力学的性質は、伸びの推定を除けば、その化学的成分のみを入力することで、ほぼ実験精度内での推定が可能であることがわかった。
Introduction
In mechanical engineering, determining the precise properties of steel is critical for ensuring structural integrity. While standards like the Japanese Industrial Standards (JIS) provide baseline requirements, the actual properties of a specific batch can vary by up to 20% due to subtle differences in composition and processing. Traditional destructive testing—such as tensile, impact, and hardness tests—is time-consuming and expensive. This paper explores the use of a neural network simulator, Neco (NEural network simulator for structure-activity COrrelation of molecules), to automate these estimations using only the chemical composition of the steel.
Methodology
The research team utilized a three-layer perceptron architecture. The input layer consisted of five neurons representing the weight percentages of Carbon (C), Manganese (Mn), Nickel (Ni), Chromium (Cr), and Molybdenum (Mo). A hidden layer of five neurons was optimized using a restructuring learning method to find the most efficient internal representation. The output layer predicted one of six mechanical properties: Yield Point, Tensile Strength, Elongation, Reduction of Area, Impact Value, or Hardness. Validation was performed using leave-one-out cross-validation on a dataset of Cr-Mo (SCM), Ni-Cr (SNC), and Ni-Cr-Mo (SNCM) steels from the Metal Data Book (1984).
Results and Discussion
The neural network demonstrated high predictive accuracy for most mechanical properties, suggesting that chemical makeup is a strong predictor of mechanical behavior. Tensile strength and yield point showed excellent correlation with experimental data. For tensile strength, the model achieved a correlation coefficient ($R^2$) of 0.9247, with a maximum relative error of only ~7%.
Hardness and reduction of area also fell within the ~20% error margin typically accepted in physical testing. However, the model struggled significantly with elongation (ductility). The correlation coefficient for elongation was a low 0.3848, with errors reaching 32%.
This discrepancy indicates that elongation is likely influenced more heavily by factors not captured in the chemical list, such as specific heat treatment temperatures or microscopic grain structures. Despite this limitation, five of the six properties were estimated with high reliability.
Conclusion
The study concludes that a simple neural network can effectively map the non-linear relationships between alloying elements and mechanical performance for most structural steels. The Neco simulator provides results in under a minute, offering a significant speedup for industrial quality control. Future research should aim to incorporate processing history into the model to improve predictions for ductility and explore more complex properties like fatigue limits.
Summary updated