Optimizing casting process using a combination of small data machine learning and phase-field simulations

Abstract It has been a challenge to employ machine learning (ML) to optimize casting processes due to the scarcity of data and difficulty in feature expansion. Here, we introduce a nearest neighbor search method to optimize the stratified random sampling in Latin hypercube sampling (LHS) and propose a new revised LHS coupled with Bayesian optimization (RLHS-BO). Using this method, we optimized the squeeze-casting process for mine fuel tank partition castings for the first time with an ultra-small dataset of 25 samples. Compared to traditional methods such as random sampling, interval sampling, orthogonal design (OD), and central composite design (CCD), our approach covers the process parameter space more, reduces the data volume by approximately 50%, and achieves process optimization beyond five factors-five levels with fewer data. Through RLHS and 6 iterations of experiments, the optimal process was identified, and the ultimate tensile strength (UTS) of partition casting under the optimal process reached 239.7 MPa, with an elongation (EL) of 12.2%, showing increases of 17.6% and 18.4% over the optimal values in the initial dataset. Finally, a combination of Shapley additive interpretation (SHAP) and phase-field method (PFM) of solidification dendrite growth was used to address the issue of weak physical interpretability in ML models.