Prediction of carbon Diffusivity in Austenite with a Neural Network model

This study develops a Neural Network (NN) model to predict carbon diffusivity (DC) in austenite under gas carburization conditions. The carburization process enhances surface hardness and wear resistance after quenching, by locally increasing the carbon content. By knowing DC, the efficiency of the carburization process can be enhanced. A fully empirical and a combined empirical-synthetic database are used for training the NN model. Synthetic data are generated using physics-based thermodynamic and mobility software, preventing data scarcity. Extensive pre-processing, including dimensionality reduction study, principal component analysis, and correlation analysis, controlling training input quality. The NN model uses supervised learning, estimating prediction and confidence intervals for DC. Results indicate that the proposed NN is able to handle experimental and synthetic data, opportunely predicting the target. This work aligns with the sustainable development goals by enhancing process efficiency and reducing energy consumption in steel manufacturing. Future developments involve extending the model to handle more complex carburization cycles and processes, such as nitriding, and refining redictive accuracy by incorporating additional data and methodologies.