Analysis of thermal conductivity of aluminum alloys by compositions and tempering process using machine learning

Thermal conductivity (TC) of commercially available Aluminum alloys is often hard to predict by machine learning (ML) algorithms due to the lack of a large dataset. The costly simulations and time-consuming experiments slow down the advancement to explore the thermal conductivity of aluminum alloy as it depends on the continuous iteration of compositional fractions and the manufacturing process. To accelerate this process, a small dataset of 271 Al alloys was organized from widely used alloys available on the market from series 1XXX to 8XXX. The dataset contains 14 alloying elements, mechanical properties, and tempering methods in manufacturing, which were preprocessed with label encoding. After performing correlation analysis among the variables, the dataset was found to be unique and reliable to train several supervised ML models. During training, 5-fold cross validation was performed to get the best set of hyperparameters employing Bayesian optimization to ensure the best performance for each model. Among them, eXtreme Gradient Boosting (XGB) predicted the TC with the highest R² value of 0.91. For this best model, feature importance analysis was performed to compare the actual metallurgical and statistical importance that affects TC of Al alloys. To improve the model accuracy further, we dropped the mechanical variables from the input variables, which depend on lab results. This additionally boosted the performance of the XGB model to an R² of 0.95. Our study demonstrates that the thermal conductivity of Al alloys can be forecasted quickly and accurately using limited datasets in ML algorithms rather than costly iterative experiments.