A New Method to Improve the Mechanism Model of Carburizing by Multivariate Linear Regression

Xinbo Gao, Xiaoxu Chen, Qiqi LiJournal of Materials Processing and Design 2022 · DOI

Abstract

Heat treatment is one of the essential technologies in industrial manufacturing, and carburization is one of the processes of heat-treatment technology. With the increase of the technological requirements of production, the optimization of carburizing technology is desired. However, most carburization mechanism models are derived from the material properties and the related physical principles, which are involved huge amounts of parameters. Some parameters (such as diffusion constant, thermal conductivity, interface transfer coefficient, etc.) are difficult to measure correctly, and this is an extremely unfavorable issue to cause the results of the mechanism model inaccurate. In this paper, a new method is raised to improve the carburization mechanism model. The function of the method is to reduce the error of the model results by combining the mechanism model and the multivariate linear regression model with a small amount of sample data. At last, the author of this paper will perform experiments to prove the correctness of the method.

1. Introduction

Carburization is a critical heat-treatment process where carbon is diffused into the surface of low-carbon steel to create a hard, wear-resistant outer layer while maintaining a tough core. To predict the depth and concentration of this carbon layer, engineers traditionally rely on "mechanism models" based on Fick’s laws of diffusion.

However, these physics-based models are notoriously difficult to calibrate. They require precise values for parameters like diffusion constants, thermal conductivity, and interface transfer coefficients. In a real-world industrial furnace, these variables are arduous to measure accurately, leading to significant discrepancies between theoretical predictions and actual results—especially at lower temperatures where material phases are less stable.

2. Methodology

The proposed approach introduces a seven-step optimization method that layers machine learning on top of classical physics. Instead of using machine learning to predict the carbon concentration from scratch, the authors use Multivariate Linear Regression (MLR) to model the error ($e_k$) between the experimental carbon concentration and the mechanism model's prediction.

Flowchart of the hybrid optimization method combining mechanism models with MLR.

As shown in the flowchart, the process is iterative. The predicted error from the MLR is added back to the mechanism model result ($R_m + e_n$). This approach is novel because it doesn't require the massive datasets typically associated with deep learning; it leverages the existing physical framework and uses MLR to "mop up" the residual inaccuracies.

3. Experimental Dataset

The researchers generated a custom dataset for this study using a professional Fengdong BBH-600-2R furnace:

  • Materials: 33 different types of low-carbon and alloy steels, including common variants like 20CrMnTi and 16MnCr5.
  • Samples: 990 total data points (30 cylindrical samples per material).
  • Conditions: Experiments were conducted at three temperatures (850°C, 920°C, and 950°C) for a duration of 8 hours with a 1.2% carbon potential.
  • Split: The database was divided into 10 sets; 9 sets were used for machine learning optimization, and 1 set was reserved for final verification.

4. Results and Discussion

The hybrid MLR-mechanism model outperformed the standalone physics model across all test temperatures. Traditional mechanism models struggle significantly with manganese-based alloys because manganese stabilizes austenite in a way that standard diffusion constants fail to capture accurately below 900°C.

Comparison between the optimized MLR results and experimental results across different temperatures and materials.

As shown in the results, the relative error was drastically reduced compared to the baseline mechanism model:

  • At 850°C: Relative error dropped from 0.28 to 0.0606.
  • At 920°C: Relative error dropped from 0.14 to 0.0456.
  • At 950°C: Relative error dropped from 0.1 to 0.0479.

This demonstrates that by modeling the residuals of a physics equation rather than the process itself, high accuracy can be achieved with relatively few samples. The model also exhibited a "self-learning" effect, where accuracy consistently improved as more experimental data cycles were ingested.

5. Conclusion

This study successfully corrected for hard-to-measure physical parameters in the carburization process, reducing prediction errors by up to 78% using a hybrid MLR-mechanism approach. While currently validated on specific industrial equipment, future work includes follow-up experiments on a wider variety of furnace types and the development of a web application for broader industrial use.

Summary updated

Abstract

Heat treatment is one of the essential technologies in industrial manufacturing, and carburization is one of the processes of heat-treatment technology. With the increase of the technological requirements of production, the optimization of carburizing technology is desired. However, most carburization mechanism models are derived from the material properties and the related physical principles, which are involved huge amounts of parameters. Some parameters (such as diffusion constant, thermal conductivity, interface transfer coefficient, etc.) are difficult to measure correctly, and this is an extremely unfavorable issue to cause the results of the mechanism model inaccurate. In this paper, a new method is raised to improve the carburization mechanism model. The function of the method is to reduce the error of the model results by combining the mechanism model and the multivariate linear regression model with a small amount of sample data. At last, the author of this paper will perform experiments to prove the correctness of the method.
BibTeX
@article{gao2022new,
  author = {Xinbo Gao and Xiaoxu Chen and Qiqi Li},
  title = {{A New Method to Improve the Mechanism Model of Carburizing by Multivariate Linear Regression}},
  journal = {Journal of Materials Processing and Design},
  year = {2022},
  doi = {10.23977/jmpd.2022.060109},
}
Figure