Owing to its low specific gravity together with high strength and high ductility, aluminum alloy die castings are being increasingly applied to the body structural parts to reduce the weight of automobiles. Generally, the tensile test is used for evaluating the properties of die castings, but this method tends to overestimate the effects of porosity defects. Focusing on the main property required of body structural parts, i.e., the ductility which guarantees that the parts will not break even if bent in case of collision, we examined the possibility to evaluate this ductility by the three-point bending test.

  In the tensile test, porosities reduce the actual cross-sectional area of the test piece, thus their macro effects on mechanical properties can be assessed. However, it is difficult to evaluate their effects on surrounding local ductility. On the other hand, it was found that with the 3-point bending test, the influence of porosities on ductility can be assessed according to their positions with respect to the pushing point, i.e., the collision site. Artificial defects were therefore formed and their effects on ductility evaluated. The results showed that porosities more than 8 mm away from the pushing position or more than 0.5 mm inside the surface of castings do not have significant impact on bending ductility.

  For the purpose of understanding the mechanism of grain refinement in cast metal by mechanical vibration, demonstration tests were conducted using a Ammonium chloride solution. Ammonium chloride solution and molds with various shape were prepared for the water model experiments. In the case of an open mold, ammonium chloride crystals crystallized on the mold wall and grew toward the center of the mold without vibration. However, when vibration was applied, countless crystals generated at the upper side of the solution immediately after pouring. In the case of a full plugged mold, no wave and convection occurred in the solution even when vibration was added, and the behavior seen was similar to the case without vibration. When only convection occurred in the full plugged mold, crystals were found to move from the mold wall into the solution. When a weir was set in the riser, it helped prevent the generation of crystals on the mold wall of the riser from migrating to the specimen. When a mold with a weir was used for Al-2%Cu alloy, crystals in the specimen did not become refined as seen in the water model experiments.

  These results demonstrate that a large amount of crystals are generated at the boundary between the surface of the molten metal and mold wall when molten metal is vibrated, and these fine crystals fill into the mold due to sedimentation or convection, thereby achieving grain refinement.

  Mg treatment is a typical graphite spheronization treatment method and is applied by many foundry factories. If the Mg concentration is insufficient, the spheronization of the graphite does not proceed, resulting in inadequate mechanical properties. Insufficient Mg concentration tends to occur when the actual Mg yield rate is lower than the predicted Mg yield rate. Therefore, the Mg yield rate is usually empirically predicted to be lower than the actual value by skilled engineers in foundry factories based on data such as S concentration of melts. Given that it is not easy for such skilled engineers to pass down tacit knowledge to others, use of artificial intelligence has recently been drawing interest as an alternative to this process which relies on the experience of skilled engineers. In this study, we proposed a machine learning model that predicts the Mg yield rate that is below the actual value by using the quantile loss as a loss function. The quantile level τ for the quantile loss was determined by performing a grid search against the accuracy metric. We statistically identified the factors that affect the Mg yield rate to convert the tacit knowledge of skilled engineers into formal knowledge by using a least absolute shrinkage and selection operator (Lasso) regression. The proposed method was implemented in multiple linear regression (MLR), gradient boosting decision tree (GBDT), and neural network (NN), and the accuracy metric was compared. The results show that GBDT has the best generalization performance. We also examined sequential batch learning, in which learning and inference are repeated alternately. The accuracy of GBDT with sequential batch learning was better than that of GBDT with batch learning.