Prediction model of dynamic fracture toughness of nickel-based alloys: combination of data-driven and multi-scale modelling

Abstract

The dynamic fracture toughness of nickel-based alloys is closely related to both the internal strengthening phases and external impact loads. However, the prediction of fracture toughness is challenging due to complex factors such as multi-scale effects, non-linearity, and strong correlations. To address these challenges, a novel prediction method for the dynamic fracture toughness of nickel-based alloys has been proposed, which integrates multi-scale numerical simulations (physical models) with machine learning techniques. This method considers both external factors (impact velocity) and internal factors (the size and distribution of strengthening phases).The study analyzes the correlation between impact velocity, the size and distribution of strengthening phases, and dynamic fracture toughness. It also proposes an optimization strategy for the hidden layer depth and the number of neurons in the neural network, resulting in an optimal network structure (8 layers with 25 neurons per layer) for predicting dynamic fracture toughness. Based on this optimized network structure, the study identifies the optimal sample size (N = 90) for ensuring a balance between computational efficiency and prediction accuracy in the fracture toughness model. Furthermore, a comparative analysis is conducted on the fitting and prediction performance of various machine learning models, including artificial neural networks (ANN), support vector machines (SVM), k-nearest neighbors (KNN), and decision trees (DT). Among these, the ANN model demonstrates the best performance, with an average absolute error of 6.73 and a coefficient of determination (R²) of 0.98. Additional validation confirms the model's excellent capability for both interpolation and extrapolation.This research successfully achieves accurate predictions of the dynamic fracture toughness of nickel-based alloys, incorporating physical information transfer. It provides new insights for the design and service performance optimization of nickel-based alloy materials.