Application of Adaptive Neuro-Fuzzy Inference System models in estimating steel hardenability

Purpose: The paper examines the efficacy of Adaptive Neuro-Fuzzy Inference Systems (ANFIS) in predicting hardenability, a crucial material property influencing the performance of steels. The study evaluates the performance of four distinct ANFIS models, each characterised by different rules (19, 22, 24, and 28), to understand how model complexity relates to predictive accuracy. The goal is to determine the optimal model structure for achieving both high accuracy and generalisation ability. Design/methodology/approach: The research utilises a dataset with a defined hardenability to train and validate the four ANFIS models. Each model’s performance is assessed using standard evaluation metrics of Root Mean Squared Error (RMSE). The analysis compares the models’ performance across training and validation datasets to identify trends in the number of rules and their impact on accuracy and generalisation. Findings: The results show a trend of increasing training accuracy with a higher rule count, suggesting a model’s enhanced ability to represent complex data relationships. However, the improvement in training accuracy is often accompanied by a reduced generalisation ability, particularly evident in the anfis-4 model. The anfis-2 model appears promising for generalisation, while the anfis-3 model demonstrates balanced performance. Research limitations/implications: The models’ applicability is limited by the specific steel grade ranges and dataset used. Future research should use a broader range of steel grades and datasets to enhance the models’ versatility and generalisability. Examining the individual rules and membership functions can provide valuable insights into the decision-making process and the factors influencing hardenability. Practical implications: The findings highlight the potential of ANFIS models as valuable tools for materials engineers and manufacturers. Accurately predicting hardenability can optimise steel production processes, thanks to reducing material waste, improving product quality, and accelerating development cycles. Such models could streamline the selection of appropriate steel grades, enabling efficient and cost-effective manufacturing. Originality/value: The study systematically investigates the effects of varying rule counts in ANFIS models for hardenability prediction. The analysis shows the trade-off between model complexity and generalisation ability, which is essential knowledge for developing robust and reliable predictive models.