Performance Analysis of Radial Basis Function Metamodels for Predictive Modelling of Laminated Composites

Area of Science:

• Materials Science
• Computational Mechanics
• Engineering

Background:

• Finite element method (FEM) provides accurate structural analysis for laminated composites but is computationally expensive.
• Metamodels offer a computationally efficient alternative to FEM once trained with sufficient data.
• Radial basis function (RBF) metamodels are explored for their potential in predictive modeling of composite structures.

Purpose of the Study:

• Investigate the efficacy of RBF metamodels for predicting laminated composite behavior.
• Evaluate the influence of different RBF basis functions on predictive accuracy.
• Assess the impact of problem dimensionality and training sample uniformity on metamodel performance.

Main Methods:

• Employed RBF metamodels for predictive modeling of laminated composites.
• Assessed various RBF basis functions and their performance.
• Studied the effect of low-dimensional (2-variable) and high-dimensional (16-variable) problems.
• Compared random sampling, Latin hypercube sampling, and Hammersley sampling for training data generation.

Main Results:

• RBF metamodels demonstrate utility in predicting laminated composite behavior.
• Metamodel performance is sensitive to the choice of RBF basis functions.
• Problem dimensionality significantly influences RBF metamodel accuracy.
• Uniformity of training samples, particularly via Latin hypercube or Hammersley sampling, enhances metamodel performance.
• Cross-validation error alone is insufficient; independent test data is vital for robust assessment.

Conclusions:

• RBF metamodels offer a viable, computationally efficient approach for analyzing laminated composites.
• Optimal metamodel performance necessitates careful selection of basis functions and uniform training data.
• Performance evaluation must include independent test datasets to avoid misleading conclusions regarding metamodel accuracy.