Quality prediction of semi-solid die casting of aluminum alloy in terms of machine learning

Researchers have developed a machine learning model to identify defective products by detecting injection pressure during the semi-solid die casting process, thereby providing a foundation for monitoring and further optimizing the manufacturing process.

Compared to traditional die casting with full liquid melt, semi-solid die casting applies a higher viscous slurry and fills the die cavity with laminar flow, thus avoiding entrapment defects and enhancing the mechanical properties. However, this technology has not yet achieved the widespread commercial application envisioned in its early stage. One of the most critical challenges is that the process window is narrower, and the process stability is poorer compared to traditional die casting. Addressing this challenge, Professor Qiang Zhu from Southern University of Science and Technology and Associate Professor Xiaogang Hu from Sun Yat-Sen University have developed a machine learning model to identify defective products through the detection of injection pressure during the semi-solid die casting process, thereby providing a foundation for monitoring and further optimizing the manufacturing process.

"Semi-solid slurry provides the ability for smooth filling of the mold cavity during die casting, but it also results in extreme sensitivity of the process to production conditions, including mold temperature and ambient temperature," explains Dr. Xiaogang Hu. "Fluctuations in these process conditions are unavoidable in actual production, leading to poor quality stability of semi-solid die castings."

"We try to establish a connection between process data and product quality using machine learning methods." Dr. Hu says, "The main challenge lies in selecting appropriate indicators to describe these process fluctuations. Based on our research experience, We chose filling pressure as a key indicator to reflect the slurry flow behavior.

The authors introduced data slicing and curve node extraction approaches based on domain knowledge in the data preparation phase. The results show that training with the filtered data yields significantly better outcomes than using the raw data directly. This indicates that the data preprocessing and feature selection methods are effective, significantly enhancing the model's predictive performance.

The authors have compared various machine learning algorithms, and the results illustrated that the multi-layer perceptron (MLP) model achieved the highest accuracy in predicting the quality of semi-solid die castings. The probability and type of defect formation can be predicted through the characteristics of the filling pressure curve. More importantly, this model has helped to reveal the mechanisms behind the formation of surface and internal defects in semi-solid die castings.

"The predictive model tells us that during the filling stage, it is not necessarily the case that the higher the solid fraction of the slurry, the smoother the filling will be," says Dr. Hu, "there is an optimal solid fraction, higher or lower than that can cause turbulence."

Professor Qiang Zhu, the leader of this research project, emphasizes that machine learning methods offer an opportunity to handle the complex nonlinear relationships among high-dimensional physical data and have been widely applied in intelligent manufacturing in recent years. Semi-solid die casting technology has not yet achieved large-scale engineering application due to its poor process stability. Utilizing the quality prediction model based on filling pressure can help us better monitor process fluctuations and provide a foundation for subsequent process interventions.

This paper "Quality prediction of semi-solid die casting of aluminum alloy in terms of machine learning" was published in Advanced Manufacturing.

Citation: Wang Z, Hu X, Li G, Xu Z, Lu H, et al. Quality prediction of semi-solid die casting of aluminum alloy in terms of machine learning. Adv. Manuf. 2024(3):0015, https://doi.org/10.55092/am20240015.