A systematic neuro-fuzzy modeling framework with application to material property prediction

Area of Science:

• Computational intelligence
• Materials science
• Machine learning

Background:

• Accurate prediction of mechanical properties in materials like hot-rolled steel is crucial for engineering applications.
• Traditional modeling approaches often struggle with the complexity and nonlinearity inherent in material science data.
• Developing efficient and automated modeling frameworks is essential for advancing materials design and manufacturing.

Purpose of the Study:

• To propose a systematic neural-fuzzy modeling framework for automated structure identification and parameter optimization.
• To apply the framework for nonlinear system identification and mechanical property prediction in hot-rolled steels.
• To validate the model's efficacy through experimental studies and demonstrate its ability to simplify complex rule-bases.

Main Methods:

• A self-organizing network for initial fuzzy model self-generation.
• Fuzzy clustering and adaptive back-propagation learning for structure identification and parameter optimization.
• Similarity analysis-based model simplification for rule-base reduction.

Main Results:

• The proposed framework successfully identified nonlinear systems and predicted mechanical properties of hot-rolled steels.
• The elicited fuzzy model, despite its simplicity (small number of rules), demonstrated a good agreement with experimental data.
• Automated processes for model generation, input selection, and simplification were achieved.

Conclusions:

• The developed neural-fuzzy modeling framework offers an efficient and automated approach to complex system modeling.
• This method provides accurate predictions of material properties, enabling potential advancements in materials science and engineering.
• The framework's ability to simplify rule-bases makes the resulting models more interpretable and computationally efficient.