Machine Learning-Driven Failure Analysis of the Effects of

1. Machine Learning-Driven Failure Analysis of the Effects of Carbides in Steels

MACHINE LEARNING-DRIVEN FAILURE
ANALYSIS OF THE EFFECTS OF CARBIDES IN
STEELS
YU HUANG, YUN HANG, DEREK O. NORTHWOOD, CHENG LIU*
COLLEGE OF MECHANICAL ENGINEERING, YANGZHOU UNIVERSITY
MECHANICAL, AUTO AND MATERIALS ENGINEERING, UNIVERSITY OF WINDSOR
*CORRESPONDING AUTHOR: LIUCH@YZU.EDU.CN

2.

PRESENTATION OUTLINE
1. Introduction & Background
• Research context and significance
• Literature review and research gaps
2. Problem Statement
• Case study: Clutch housing roller die failure
3. Methodology
• Data collection and feature extraction
• Machine learning algorithms and evaluation
4. Results & Discussion
• Model performance and feature importance
• Microstructural validation
5. Conclusions & Future Directions

3. 1. Introduction: Machine Learning in Materials Science

1. INTRODUCTION: MACHINE LEARNING IN
MATERIALS SCIENCE
• Machine Learning (ML) is a data analysis method used to uncover intrinsic relationships within complex datasets.
• Current Applications:
• • Property Prediction: Creep fracture life in ODS steels [4]
• • Mechanical Properties: Tensile strength, yield strength, elongation [5]
• • Material Design: Efficient screening of new compositions [6]
• Research Gap:
• Quantitative analysis of microstructure-property relationships applied to engineering failure analysis remains
underdeveloped.

4. Microstructure-Property Correlation

MICROSTRUCTURE-PROPERTY CORRELATION
Microstructure is the fundamental factor determining macroscopic properties [7].
• Critical Carbide Features:
• Size: Influences strength and toughness
• Morphology: Affects stress concentration
• Distribution: Impacts fatigue performance
Previous Work:
• Li et al. [8]: EBSD characterization for low-alloy steels
• Gorynski et al. [9]: Grain shape and orientation quantification
• Gap: No quantitative correlation between carbide features and failure analysis.

5. 2. Case Study: Roller Die Failure

2. CASE STUDY: ROLLER DIE FAILURE
• Component: Roller die for automotive clutch housings
• Service Life Analysis:
• • Expected Life: ~80,000 units
• • Actual Life (Failed): 29,790 units (37.2%)
• • Reference (High-Life): 72,558 units
• • Performance Gap: 2.43? difference
Key Metrics
80,000 Expected
29,790 Actual
37.2% Achievement
• Objective: Establish quantitative analytical chain linking carbide features to failure mechanisms.

6. 3. Methodology Framework

3. METHODOLOGY FRAMEWORK
Multi-Parameter Quantitative Analysis using Image-Pro Plus (IPP)
• Step 1: Data Collection
• 88 steel grades from literature and commercial sources
• SEM characterization and IPP measurements
Step 2: Feature Extraction
• 40 feature vectors including Radius-min, Center-X/Y, Aspect
Step 3: ML Modeling
• Algorithm selection and hyperparameter optimization
• • 10-fold cross-validation

7. Dataset Construction & Features

DATASET CONSTRUCTION & FEATURES
Dataset Statistics:
• 88 steel grades, >3,000 measurements
• 40 input features, 3 mechanical properties (YS, UTS, EL)
Preprocessing:
• Min-Max Normalization: y = (xi - xmin) / (xmax - xmin)
• PCA for feature selection and dimensionality reduction
Key Features:
• Radius-min: Min distance between centroid and outline
• Center-X/Y: Coordinates of object's centroid
• Aspect: Ratio of major to minor axis
• Dendritic Length: Total length of dendrites

8. ML Algorithms & Evaluation Metrics

ML ALGORITHMS & EVALUATION METRICS
Algorithms Evaluated:
• Random Forest (RF), Gradient Boosting Decision Tree (GBDT)
• XGBoost, AdaBoost, Extremely Randomized Trees (ETR)
• Decision Tree, K-Nearest Neighbors (KNN)
Top Performers:
• AdaBoost: Best for Elongation (EL)
• GBDT: Best for Yield Strength (YS) & UTS
Evaluation Metrics:
• R? (Coefficient of Determination)
• RMSE (Root Mean Square Error)
• MAE (Mean Absolute Error)
• PCC (Pearson Correlation Coefficient)

9. Feature Selection & Correlation

FEATURE SELECTION & CORRELATION
• Pearson Correlation Analysis:
• • High linear correlation between carbide roundness and dendrite distribution
• • Features with high correlation can substitute for one another [10]
• Principal Component Analysis (PCA) [13]:
• • Reduces feature count and computational complexity
• • Preserves core information
• • Identifies key features contributing to output
• • Eliminates redundant features
• Strategy: 80% Training / 20% Testing with 10-fold cross-validation

10. Results: Elongation (EL) Prediction

RESULTS: ELONGATION (EL) PREDICTION
AdaBoost Model - Superior Performance
Metrics:
• R? = 0.95 (Excellent accuracy)
• RMSE = 3.92
• MAE = 3.22
Characteristics:
• High accuracy and stability
• Low bias and low variance
• Effective for small-to-medium datasets
• Successfully captures difficult-to-predict data points
Top Feature: Radius-min (strongest positive influence)
0.95
3.92
3.22
R?
RMSE
MAE

11. Results: YS & UTS Prediction

RESULTS: YS & UTS PREDICTION
GBDT Model Performance
• Yield Strength (YS):
• R? = 0.81 | RMSE = 146.83 | MAE = 121.72
• Reliable explanation of majority variance
Ultimate Tensile Strength (UTS):
• R? = 0.68 | RMSE = 148.61 | MAE = 107.36
• Prediction biases in high/low intensity regions
Note: Lower UTS scores indicate carbide features alone are insufficient.
• UTS depends on microstructural characteristics AND process parameters.

12. Feature Importance Analysis

FEATURE IMPORTANCE ANALYSIS
Critical Parameter: Radius-min
Definition: Minimum distance between carbide's centroid and outline
Ranking by Property:
• EL: 1. Radius-min, 2. Center-X, 3. Center-Y Mass
• YS: 1. Center-Y, 2. Center-Y Mass
• UTS: 1. Radius-min, 2. Dendritic Length
Physical Significance:
• Small Radius-min ? Flat/elongated shapes
• Large Radius-min ? Circular/spherical shapes
• Captures morphological differences, not just size
Conclusion: Morphological effect outweighs size effect

13. Microstructural Validation (SEM)

MICROSTRUCTURAL VALIDATION (SEM)
Failed Die (29,790 cycles):
• Large carbide particles
• • Flat/elongated shapes (small Radius-min)
• Coarse eutectic carbides
High-Life Die (72,558 cycles):
• Uniformly distributed carbides
• Predominantly circular (large Radius-min)
• Consistent size (2.43? longer life)
Validation: ML predictions confirmed experimentally.
• Flat/elongated carbides reduce EL and UTS as predicted.

14. Discussion: Failure Mechanism

DISCUSSION: FAILURE MECHANISM
Key Insight: Morphology > Size
Failure Mechanism:
1. Irregular carbides induce local stress concentration
2. Preferred nucleation sites for microcracks
3. Accelerated damage accumulation
4. Premature mold failure
Engineering Implication:
Control of carbide morphology (maximizing Radius-min) is crucial
for optimizing properties and preventing premature failure.
Model Limitations:
UTS prediction requires additional process parameters for improvement.

15. Conclusions & Future Directions

CONCLUSIONS & FUTURE DIRECTIONS
Main Conclusions:
• 88 steel samples, 3,000+ measurements analyzed
• AdaBoost: R? = 0.95 for EL (high accuracy)
• GBDT: R? = 0.81 (YS), R? = 0.68 (UTS)
• Radius-min: Most influential for failure mechanisms
• Novel IPP-based method validated
Future Research:
• Incorporate steel preparation process parameters
• Improve UTS prediction accuracy
• Extend to other material systems
Thank You for Your Attention
Contact: liuch@yzu.edu.cn