From Root Cause Analysis to Predictive Quality: A Dual AI Approach to Slash Die Casting Scrap Rates

This technical brief is based on the academic paper "TOWARDS AN AI-Driven Smart Manufacturing of Diecastings" by F. Liu, S. Wang, X. Liu, T. Zhang, B. Yang, Q. Han, D. Yang, and Corey Vian, published in NADCA Transactions 2018. It is summarized and analyzed for HPDC professionals by the experts at CASTMAN.

Keywords

• Primary Keyword: AI in Die Casting
• Secondary Keywords: Smart Manufacturing, Die Casting Process Optimization, Machine Learning for Manufacturing, Scrap Rate Reduction, XGBoost, Generalized Linear Model (GLM), Process Parameter Analysis

Executive Summary

• The Challenge: Die casting operations generate vast amounts of process data, but manufacturers struggle to use this data to effectively understand the root causes of defects and reduce high scrap rates.
• The Method: Researchers applied a dual approach to a large dataset (345,465 parts) from a real-world casting plant: (1) Statistical Generalized Linear Models (GLM) to explain the cause-and-effect relationships between process variables and part quality, and (2) advanced machine learning models (DNN and XGBoost) to predict PASS/FAIL outcomes.
• The Key Breakthrough: The statistical GLM approach successfully identified specific process parameters and their optimal operating windows to minimize defects. For predictive tasks, the XGBoost model proved superior to Deep Neural Networks (DNN), achieving 94% accuracy in classifying part quality.
• The Bottom Line: This study demonstrates that AI provides two powerful, practical pathways for die casters: an explanatory path to guide process engineers in optimizing machine settings, and a predictive path to enable automated, data-driven quality control.

The Challenge: Why This Research Matters for HPDC Professionals

In modern manufacturing, data is collected at every stage of production. For die casters, this includes everything from alloy composition and shot sleeve parameters to die temperatures and X-ray inspection results. While this data holds the key to process improvement, analyzing it to find actionable insights is a monumental task. As stated in the paper's introduction, the goal is to "apply machine learning techniques to analyze the collected data and to pursue smart manufacturing."

This research directly addresses a core pain point for the industry: how to move from simply collecting data to using it to make intelligent decisions that reduce scrap and improve consistency. The study tackles two fundamental questions that every plant manager and process engineer faces:

1. Why are we producing scrap? Which of the 91 different process factors are actually causing defects?
2. Can we predict a bad part before it gets to final inspection? Is the data we collect sufficient to automatically classify a part as "good" or "bad"?

The Approach: Unpacking the Methodology

The researchers analyzed a dataset from a FCA casting plant producing crossmember castings, which included 310,221 "PASS" parts and 35,244 "FAIL" parts. To answer the questions above, they pursued three distinct experiments:

1. Experiment 1: Statistical Analysis (GLM): They used Generalized Linear Models to investigate the relationship between 91 process variables and the probability of a part passing inspection. This traditional statistical method is excellent for discovering and explaining the factors that cause higher scrap rates.
2. Experiment 2: Deep Neural Networks (DNN): They built and trained several Deep Neural Network models, a popular AI technique, to predict whether a part would pass or fail. This experiment tested the power of complex, brain-like models for automated classification. To handle the imbalanced dataset (many more good parts than bad), they used the SMOTE over-sampling technique.
3. Experiment 3: XGBoost: Recognizing the challenges of tuning DNNs, they turned to XGBoost, a powerful and efficient tree-boosting algorithm. This approach also aimed to predict PASS/FAIL outcomes but is often faster to train and highly effective on the structured, tabular data common in manufacturing.

The Breakthrough: Key Findings & Data

The study yielded clear, practical results that highlight the distinct strengths of different AI techniques.

• Finding 1: Statistical Models Provide a Clear "Why" (GLM): The GLM analysis successfully identified which process variables had the most significant impact on part quality. It distinguished between variables that needed to be maximized, minimized, or kept within a specific optimal range. For example, Table 1 shows variables like Copper (Cu) content and Low Pressure Water Temperature (LPWT) have an optimal "maximizer" value, while Table 2 shows variables like Silicon (Si) content and Plunger Speed (PS) have a "minimizer" value that should be avoided. This provides a direct, data-driven guide for process optimization.
• Finding 2: Deep Neural Networks are Powerful but Complex (DNN): The DNN models achieved a respectable, but limited, final testing accuracy of 90.8% (Figure 7). The authors note that "tuning DNN is often long and a tedious job," and despite trying various architectures (Figure 8), they could not easily push the performance higher. This reflects a common industry experience where the complexity of DNNs can be a barrier to rapid deployment.
• Finding 3: XGBoost Excels at Prediction: The XGBoost model outperformed the DNNs, achieving a final accuracy of 94% after careful parameter tuning (Figure 10 and Figure 11). The paper concludes that "XGBoost works very well on the structured data and should be considered by the practitioners for classification problems in diecasting plants." This positions XGBoost as a more practical and immediately deployable tool for predictive quality tasks.

Practical Implications for Your HPDC Operations

This research is not just academic; it offers a concrete framework for implementing AI-driven improvements in a real-world die casting facility.

• For Process Engineers: The findings from the GLM analysis (Tables 1 & 2) serve as a direct playbook for process optimization. Instead of relying on trial-and-error, you can use this data to adjust critical parameters like alloy composition (Cu, Si, Mn), shot sleeve settings (Biscuit Length, Plunger Speed), and die cooling (water temperatures and vacuum levels) to directly influence and improve the pass rate. The joint effects shown in Figures 1-4 illustrate how changing two parameters simultaneously can create optimal (or suboptimal) conditions.
• For Quality Control: The 94% accuracy achieved by the XGBoost model demonstrates the feasibility of an automated quality prediction system. Such a system could flag potentially defective parts in real-time based on their process data, allowing for immediate intervention and reducing the workload on downstream manual or X-ray inspections. This moves quality control from a reactive to a proactive function.
• For Die Design and Management: The study's insights into parameters like Biscuit Length (BL), Cycle Time (CY), and Intensification Pressure (IP) provide data-driven evidence to inform die design and process management strategies. For example, the finding that a shorter cycle time improves the pass rate has direct implications for how processes are structured to maximize both quality and throughput.

Paper Details

TOWARDS AN AI-Driven Smart Manufacturing of Diecastings

1. Overview:

• Title: TOWARDS AN AI-Driven Smart Manufacturing of Diecastings
• Author: F. Liu, S. Wang, X. Liu, T. Zhang, B. Yang, Q. Han, D. Yang, and Corey Vian
• Year of publication: 2018
• Journal/academic society of publication: NADCA Transactions 2018, T18-072
• Keywords: AI, Smart Manufacturing, Die Casting, Machine Learning, Scrap Rate, Process Optimization

2. Abstract:

This paper describes our initial effort in applying supervised machine learning approaches to analyze data collected at a FCA casting plant during the production of crossmember castings. The data contain results of X-ray inspection on castings and processing variables such as alloy compositions, cooling conditions of the shot tooling and die, and the operation parameters of the diecasting machine. In addition to the conventional statistical machine learning approaches, such as polynomial regressions, logistic regressions and decision trees, deep neural networks were also used to investigate the relationship between the scrap rates of the crossmember casting and the operation variables of the casting process. Various data science techniques, such data cleaning, data normalization and data augmentation were also applied to further boost the validity and accuracy of the results. The finding of this work shows that machine learning approaches have an excellent potential to reduce scrap rates of the castings. The work also exemplifies how state-of-art artificial intelligence techniques can be applied in die casting to enable smart manufacturing.

3. Introduction:

As manufacturing processes become more instrumented, vast amounts of data are collected automatically. This paper explores the application of data science and artificial intelligence to analyze data from a FCA die casting plant producing crossmember castings. The dataset contains 91 process factors and results for 345,465 parts, each labeled as "PASS" or "FAIL". The research aims to conduct two types of studies: first, to use statistical analysis to discover the root-cause factors for high scrap rates, and second, to use modern machine learning algorithms to predict if a part will pass inspection, potentially automating the process.

4. Summary of the study:

Background of the research topic:

The research is set against the backdrop of Industry 4.0 and smart manufacturing, where data is a key asset. In die casting, numerous variables related to alloy chemistry, machine operations, and die conditions interact in complex ways to determine the final quality of a part. Understanding these interactions is critical for reducing defects like porosity and improving overall yield.

Status of previous research:

The paper builds on previous work in statistical analysis of manufacturing data and the application of machine learning algorithms like neural networks. It specifically references prior classification of process parameters into three categories: alloy composition, shot sleeve processes, and die/mold processes [2].

Purpose of the study:

The study has a dual purpose:

1. To identify the most important process variables that influence the pass/fail rate of die castings and determine their optimal settings to improve quality.
2. To evaluate the effectiveness of modern AI techniques (Deep Neural Networks and XGBoost) in building a predictive model capable of classifying parts as good or bad, with the goal of potentially automating inspection.

Core study:

The core of the study is divided into three experiments performed on a large, real-world dataset:

1. Experiment 1 (Statistical Methods): Application of Generalized Linear Models (GLM) to find statistically significant relationships between process variables and pass probability.
2. Experiment 2 (Deep Neural Networks): Construction and tuning of DNNs to solve the binary classification problem (PASS/FAIL).
3. Experiment 3 (XGBoost): Application of the XGBoost algorithm, a gradient boosting method, to the same classification problem to compare its performance against DNNs.

5. Research Methodology

Research Design:

The study was designed as a comparative analysis of different modeling techniques on a single, large dataset.

• For the statistical analysis (Experiment 1), logistic, probit, and cloglog models were fitted to find linear and quadratic relationships between variables and the outcome.
• For the machine learning predictions (Experiments 2 & 3), the dataset was split into training and testing sets. The models were trained on the training data and their performance was evaluated on the unseen test data.

Data Collection and Analysis Methods:

The data was collected from a FCA casting plant and stored in a structured CSV format. It included 91 columns (factors) and 345,465 rows (parts). The dataset was highly imbalanced (310,221 PASS vs. 35,244 FAIL). To address this, key data science techniques were used:

• Data Cleaning and Normalization: Standard pre-processing steps to ensure data quality.
• Feature Engineering: Time-related features like "year", "month", "day", and "hour" were extracted from timestamps for the XGBoost model.
• Over-sampling (SMOTE): In the DNN experiment, the Synthetic Minority Over-sampling Technique (SMOTE) was used to balance the training data, preventing the model from being biased towards the majority "PASS" class.

Research Topics and Scope:

The research covers three main topics:

1. Alloy Compositions: The effect of elements like Si, Cu, Zn, Mn, etc.
2. Process Parameters (Shot Sleeve): The effect of variables like Cycle Time (CY), Biscuit Length (BL), Plunger Speed (PS), and Intensification Pressure (IP).
3. Process Parameters (Die/Mold): The effect of variables like vacuum pressure (MPV1, MPV2) and water cooling temperatures/pressures (LPWT, IWT, etc.).

6. Key Results:

Key Results:

• Statistical GLM: The models successfully identified numerous variables with statistically significant linear and quadratic effects on the pass rate. Table 1 lists variables where a specific optimal value (maximizer) exists to achieve the highest pass probability. Table 2 lists variables where a "worst" value (minimizer) exists that should be avoided. The analysis also visualized the joint effects of two variables, showing the existence of local maximizers (Figure 1), local minimizers (Figure 2), and saddle points (Figure 3).
• Deep Neural Network (DNN): The DNN models showed a rapid initial improvement in training but struggled to push testing accuracy beyond 90.8% (Figure 7). The performance was sensitive to the number of neurons, but even the best-performing models hit a similar ceiling (Figure 8).
• XGBoost: The XGBoost model demonstrated superior predictive power. After tuning hyperparameters like the number of trees (n_estimators) and tree depth (max_depth), the model achieved a cross-validation accuracy of 94% on the test set, showing a clear advantage over the DNN models in this specific application.

Figure Name List:

• Figure 3- Probability of PASS as functions of two explanatory variables when one quadratic term is negative and the other is positive, where a local saddle point appears.
• Figure 4- Probability of PASS as functions of two explanatory variables when only one quadratic term is significant.
• Figure 5- The basic architecture of Multi-Layer Perceptron
• Figure 6- The training loss and testing loss of deep neural network
• Figure 7- The training accuracy and testing accuracy of DNN
• Figure 8- The test accuracy of DNNs with different neuron nodes
• Figure 9- Weak trees form a strong model
• Figure 10- Training/cv accuracy for n_estimators
• Figure 11- Training/cv accuracy for colsample and subsample

7. Conclusion:

The study concludes that traditional statistical analysis, when used correctly, is a powerful tool for explaining cause-and-effect relationships in a production environment, providing actionable guidance for process improvement. While Deep Neural Networks are promising, they can be complex and time-consuming to tune. For structured data classification problems in die casting, XGBoost offers a highly effective and more straightforward alternative. The work demonstrates that AI techniques have excellent potential to reduce scrap rates and enable smart manufacturing. Future work could be improved by collecting more data (e.g., environmental data, X-ray images for use with Convolutional Neural Networks).

8. References:

1. Faraway, J.J. Extending the Linear Model with R. CRC Press, Boca Raton, Florid, (2016)
2. Han, Q., McClure, D., Wood, D., and Yang, D. Statistical Analysis of the Effect of Operational Parameters on the Scrap Rates of Crossmember Casting, North American Die Casting Engineer, 38-43, November 2017.
3. Felmings, M.C., Solidification Processing, McGraw-Hill, USA (1974).
4. Joarder Kamruzzaman, Rezaul K.Begg, Ruhul and A.Sarker, Artificial neural networks in finance and manufacturing, Idea Group Publishing, Hershey, PA, (2006)
5. Cihan H.Dagli, Artificial neural networks for intelligent manufacturing, 1st ed., Springer Science+Business Media Dordrecht, (1994)
6. Huang, S.H., Zhang, H.C., "Artificial Neural Networks in Manufacturing: Concepts, Applications, and Perspectives," IEEE Transactions on components, Packaging and Manufacturing Technology-Part A 17(2), 212-227 (1994)
7. Kuo, R.J., Wang, Y.C., Tien, F.C., "Integration of artificial neural network and MADA methods for green supplier selection," Journal of Cleaner Production, 18(12), 1161-1170 (2010)
8. Adam, A., Chew, L.C., Shapiai, M.I., Jau, L.W., Ibrahim, Z., & Khalid, M., “A Hybrid Artificial Neural Network-Naive Bayes for Solving Imbalanced Dataset Problems in Semiconductor Manufacturing Test Process," Hybrid Intelligent Systems (HIS), 2011 11th International Conference on, pp. 133-138 (2011)
9. Haibo He, Edwardo A.Garcia, "Learning from Imbalanced Data," IEEE Transactions on Knowledge and Data Engineering-Page 1263-1284, 1041-4347(2009)
10. Chawla, N.V., Japkowicz, N., Kolcz, A., "Editorial: Special Issue on Learning from Imbalanced Data Sets," SIGKDD Explorations Special Issue on Learning from Imbalanced Datasets 6(1), 1-6(2004)
11. Jeatrakul, P., Wong, K.W., Fung, C.C., “Classification of Imbalanced Data by Combining the Complementary Neural Network and SMOTE Algorithm”, ICONIP 2010.LNCS, vol.6444, pp.152-159.Springer, Heidelberg(2010)
12. Clevert, D.-A., Unterthiner, T., Hochreiter, S., "Fast and Accurate Deep Network Learning by Exponential Linear Units (ELUs)", arXiv preprint arXiv:1511.07289(2015)
13. Zheng J, "Cost-sensitive boosting neural networks for software defect prediction", Expert Syst Appl 37(6):4537-4543, (2010)
14. Aurélien Géron, Hands-On.Machine.Learning.with.Scikit-Learn.and.TensorFlow, 1st ed., Published by O'Reilly Media, Inc., 1005 Cravenstein Highway North, Sebastopol, CA 95472(2017)
15. Gareth James, Daniela Witten, Trevor Hastie, Robert Tibshirani, An Introduction of Statistical Learning, Springer Science+Business Media New York, Vol 112, NY(2013)
16. Tianqi Chen, Carlos Guestrin. “XGBoost: A Scalable Tree Boosting System". In 22nd SIGKDD Conference on Knowledge Discovery and Data Mining, 2016.
17. Jerome Friedman. Greedy function approximation: A gradient boosting machine. The Annals of Statistics, 2001.
18. Wei-Yin Loh. Fifty Years of Classification and Regression Trees. International Statistical Review (2014)
19. Sebastian Ruder. An overview of gradient descent optimization algorithms. arXiv (2016).
20. Mitchell R, Frank E. Accelerating the XGBoost algorithm using GPU computing. PeerJ Computer Science (2017).

Conclusion & Next Steps

This research provides a valuable roadmap for enhancing process control and quality assurance in die casting. The findings offer a clear, data-driven path toward improving quality, reducing defects, and optimizing production.

CASTMAN is committed to applying cutting-edge industry research to solve our customers’ most challenging technical problems. If the problem discussed in this white paper aligns with your research goals, please contact our engineering team to discuss how we can help you apply these advanced principles to your research.

Expert Q&A:

• Q1: The paper mentions two AI approaches: statistical models (GLM) and predictive models (XGBoost). Which one is better for my factory? A: It depends on your goal. The paper, "TOWARDS AN AI-Driven Smart Manufacturing of Diecastings," suggests a dual approach. If your goal is to understand why defects are happening and guide your engineers on how to adjust machine settings, the statistical GLM approach is ideal because it explains cause-and-effect. If your goal is to automatically predict if a part is good or bad in real-time to improve quality control, the XGBoost model is the better choice, as it demonstrated 94% predictive accuracy.
• Q2: Which specific process parameters were found to be most critical for reducing defects? A: The study identified several critical parameters. According to Table 1, variables like Copper (Cu) content, Intensification Stroke (IS), and several water temperature settings (WT4, WT13, LPWT) have an optimal "maximizer" value that improves the pass rate. Conversely, Table 2 shows that variables like Silicon (Si) content, Plunger Speed (PS), and Biscuit Length (BL) have a "minimizer" value, or a point of worst performance, that should be avoided. The optimal values are calculated and listed in the tables, providing a direct guide for process adjustment.
• Q3: How accurate was the AI in predicting whether a part would pass or fail inspection? A: The paper tested two types of predictive models. The Deep Neural Network (DNN) models achieved a maximum testing accuracy of 90.8%, as shown in Figure 7. However, the XGBoost model proved to be more effective, achieving a final accuracy of 94% after tuning, making it the more powerful predictive tool for this specific dataset.
• Q4: The study mentions "local maximizers" and "minimizers" for process variables. What does this mean in practice for an engineer? A: This is a key insight from the statistical analysis in "TOWARDS AN AI-Driven Smart Manufacturing of Diecastings." A "local maximizer" (Figure 1) represents a combination of settings for two variables (e.g., LPWT and MPV1) that produces the highest probability of a "PASS" part. An engineer should aim for these settings. A "local minimizer" (Figure 2) represents a combination of settings (e.g., Biscuit Length and Plunger Speed) that produces the lowest probability of a "PASS" part, or the highest scrap rate. An engineer should actively avoid operating in this region. These visualizations provide a practical map for process control.

Copyright

• This material is an analysis of the paper "TOWARDS AN AI-Driven Smart Manufacturing of Diecastings" by F. Liu, et al.
• Source of the paper: https://www.researchgate.net/publication/366674062
• This material is for informational purposes only. Unauthorized commercial use is prohibited.
• Copyright © 2025 CASTMAN. All rights reserved.