Combination of thermodynamic knowledge and multilayer feedforward neural networks for accurate prediction of MS temperature in steels

Abstract

The combination of multilayer feedforward neural networks (MLFFNN) and thermodynamic knowledge has excellent advantages to study the complicated phenomena in material science. In the present study, a thermodynamic knowledge-based MLFFNN has been developed to predict the martensite start temperature (MS) of steels that accounts for the variations in critical driving force (ΔGc) and austenitization temperature (Tγ). This is achieved by integrating two crudely estimated properties (ΔGc and Tγ) in the feature space. Instead of using the original dataset directly, the feature space was reshaped by kernel principal component analysis (KPCA). The genetic algorithm was implemented to find a suitable hyperparameters set that was able to induce a model with high predictive capability. The resulting neural network performs well in the present dataset, and the RMSE in the unseen datasets is 21.52 K. Benchmarking of the MLFFNN predictions against JmatPro-V8 calculations also shows a significant improvement in predictive accuracy. Results indicate KPCA and the integration of crude estimation property (CEP) boost the predictive accuracy of MLFFNN. Besides, the present study also demonstrates the CEP strategy is general enough to be employed in several well-known machine learning methods including support vector regression, decision tree and Gaussian process regression.

Keywords

• Multilayer feedforward neural networks
• MS temperature
• Kernel principal component analysis
• Crude estimation property
• Thermodynamic