Prediction of Creep Fracture Life of IN718 Superalloy Based on Artificial Neural Networks: Application of Machine Learning in Materials Science

This article introduces a research project that uses artificial neural networks (ANN) to predict the creep fracture life of nickel-based superalloy IN718, aiming to overcome the limitations of traditional extrapolation methods in terms of accuracy and mechanism, and provide new ideas for material performance prediction in aerospace and other fields. The study demonstrates the application value of machine learning in materials science, using a technology stack including TensorFlow to build models and verify their effectiveness.

In aerospace, energy power generation and other fields, IN718 superalloy is widely used due to its excellent performance, but predicting creep fracture life under long-term high temperature and pressure is a core challenge. Traditional extrapolation methods have two major limitations: first, accuracy is limited, and errors amplify exponentially over time; second, lack of mechanistic support, making it difficult to explain the correlation between microstructural evolution and macroscopic failure, and deviations may lead to catastrophic consequences.

This project builds an ANN-based prediction framework, using its nonlinear mapping ability to automatically learn material response rules without preset physical models. The technology stack includes TensorFlow/Keras, Scikit-learn, and XGBoost. The unique value of deep learning is verified by comparing ANN with traditional models (linear regression, decision trees, etc.).

High-quality data is key. Steps include: 1. Correlation analysis and heatmap visualization to identify key parameters; 2. ExtraTrees regressor to calculate feature importance for variable screening; 3. Correlation coefficient threshold (0.9) to filter multicollinearity; 4. Standardization of numerical features; 5. Split the dataset into training/test sets in an 80:20 ratio, and set a random seed to ensure reproducibility.

The neural network uses a sequential architecture: the input layer receives standardized features, the hidden layers use ReLU activation, and the output layer performs regression to predict life. A two-layer hidden structure (neurons decreasing layer by layer) is designed to avoid overfitting. Hyperparameters (number of neurons, learning rate, batch size, number of training epochs) are optimized via RandomizedSearchCV, and the optimal parameters (maximizing R²) are selected using 3-fold cross-validation.

The optimized ANN model achieves R²=0.98 on the training set and R²=0.97 on the test set, showing excellent generalization ability. Predicted values and measured values are closely distributed around the diagonal line, with uniform and unbiased errors. Comparisons show that ensemble methods like XGBoost are competitive, but ANN is slightly better at capturing nonlinear relationships. SHAP values are used to improve model interpretability.

Research significance: 1. Aerospace engine manufacturers can optimize maintenance cycles; 2. Material research and development can guide new alloy design; 3. Nuclear power plants can establish risk equipment replacement strategies. Future directions: Integrate physical prior knowledge (such as PINN), use transfer learning to share alloy system experience, and promote the application of data-driven methods in materials science.