Regularized radial basis function for mixed-mode fracture parameters prediction using artificial neural networks

Traditional artificial neural networks (ANNs) in fracture mechanics often suffer from limitations such as inconsistent hyperparameter optimization, reliance on default settings without sensitivity analyses, and difficulties in ensuring generalization and reproducibility across datasets. Accurate prediction of mixed-mode fracture parameters stress intensity factors ( K I , K II ) and the T-stress is critical for structural integrity assessment. This investigation systematically evaluates ANN configurations for predicting mixed-mode fracture parameters in PMMA short beam bend specimens. It identifies the Bayesian Regularized Radial Basis Function Network (BRBFN) as a globally optimal surrogate model through an exhaustive evaluation of 6,552 unique ANN configurations, incorporating 13 training algorithms, 12 activation functions, and 42 network topologies. The BRBFN achieves exceptional accuracy on validation data ( R 2 > 0 . 997 , MSE < 0 . 033 ) with training completed in under one second. A comparative analysis against advanced deep learning architectures Convolutional Neural Networks (CNN), Long Short-Term Memory networks (LSTM), and Gated Recurrent Units (GRU) demonstrates that the BRBFN significantly outperforms these complex models in generalization on unseen data (MSE improvement of 27%–35%), while requiring two orders of magnitude less training time. A six-metric evaluation framework (MSE, RMSE, MAE, IAE, ITAE, R 2 ) confirms the robustness of this configuration across complementary performance criteria. This establishes a reproducible and efficient methodology for deploying reliable ANNs in fracture mechanics, minimizing overfitting risks in limited-data regimes. Validation against experimental fracture initiation angles confirms practical applicability, with predictions closely mirroring the generalized maximum tangential stress criterion.