⚡ electrical engineering

Machine Learning Assisted Prediction and Experimental Comparison of Punch Wear in Blanking of AISI D2 and AISI D3 Cold Work Tool Steels

This study combines experimental trials on a 500-ton hydraulic press with machine learning modeling to demonstrate that AISI D3 tool steel offers superior wear resistance over AISI D2 for blanking operations and that Random Forest algorithms can accurately predict punch wear evolution based on material properties and stroke count.

Original authors: Hasan Öktem, Gürkan Soy, Naresh Kumar Konada

Published 2026-10-09

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Hasan Öktem, Gürkan Soy, Naresh Kumar Konada

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the factories that build cars, appliances, and buildings, a massive machine repeatedly slams a sharp metal punch through sheets of steel to cut out shapes. This process, known as blanking, is the workhorse of modern manufacturing, creating the precise parts that hold our world together. However, the sharp edge of the punch is not indestructible. With every strike, friction and pressure wear it down, causing the edge to become slightly rounded. As the edge dulls, the quality of the cut deteriorates, leaving rough edges or burrs on the finished parts. For decades, factory managers have managed this wear by guessing when to stop and sharpen the tool, often changing it too early and wasting money, or too late and risking a broken machine. The challenge lies in predicting exactly how long a tool will last before it fails, a task that depends on the specific metal the tool is made of and how hard it is pushed.

A team of researchers set out to solve this puzzle by comparing two common types of steel used for these punches: AISI D2 and AISI D3. They wanted to see which metal could withstand the punishment of cutting through one-millimeter-thick low-carbon steel sheets for longer. To find out, they ran a massive experiment on a 500-ton hydraulic press. They installed punches made of each steel type and let them cut continuously, stopping only at specific intervals to measure the damage. They tested them from 10,000 strokes all the way up to 80,000 strokes, all while keeping the conditions dry and identical, with no oil or lubricant to interfere. This dry environment is common in many industries where oil would contaminate the final product, but it makes the job much harder for the tool.

The results revealed a clear winner in the early stages of the test. The punch made from AISI D3 steel, which is harder and contains more carbon, held its sharp edge significantly better than the AISI D2 punch. At the 10,000-stroke mark, the D3 punch had worn down so much less that its volume loss was nearly five times smaller than that of the D2 punch. Microscopic examination showed why: the D2 punch began to suffer from a specific type of damage where tiny pieces of the workpiece steel stuck to the tool's edge, forming a rough, jagged buildup that accelerated wear. The D3 punch, being harder, resisted this sticking and wore down smoothly and evenly, like a stone slowly smoothing in a river, without any of that messy buildup.

However, the story changed as the number of cuts increased. As the experiment pushed toward 80,000 strokes, the advantage of the harder D3 steel began to shrink. By the end of the test, the difference in wear between the two materials had narrowed dramatically, with the D3 punch showing only a slight edge over the D2. This happened because the D2 punch eventually developed a layer of built-up material that temporarily masked its wear, making the two tools look more similar in their final state. The researchers found that while the D3 tool remained cleaner and sharper throughout the entire process, the sheer volume of material lost by the D2 tool caught up significantly over time.

To turn these physical observations into a useful tool for the future, the researchers fed their data into three different computer models designed to learn from patterns. They used a method called machine learning, where a computer program studies the relationship between the number of cuts, the hardness of the steel, and the resulting wear to predict what will happen next. Among the three models they tested, one called Random Forest proved to be the most accurate. It learned the complex, non-linear way the tools wore down with such precision that it could predict the wear volume with nearly 99% accuracy. The computer identified that the number of strokes was the single most important factor in predicting wear, followed closely by the hardness of the steel.

This study offers a practical path forward for industries that rely on these cutting tools. By combining real-world testing with smart computer models, factory managers can now move away from guessing and toward precise planning. They can choose the right steel for their specific production needs, knowing that for shorter runs, the harder D3 steel offers a clear benefit, while for very long runs, the difference becomes less critical. The research confirms that while material science provides the foundation, the ability to predict wear through data allows for smarter, more efficient manufacturing, ensuring that the machines keeping our world moving run smoothly for as long as possible.

Technical Summary: Machine Learning Assisted Prediction and Experimental Comparison of Punch Wear in Blanking of AISI D2 and AISI D3 Cold Work Tool Steels

Problem Statement
The blanking process is critical for manufacturing components in the automotive, electronics, and construction sectors. However, tool wear management remains largely empirical, relying on fixed, conservative re-grinding intervals derived from trial-and-error rather than predictive models. This approach often leads to premature tool removal or unplanned downtime due to catastrophic failure. While cold-work tool steels AISI D2 and AISI D3 are dominant materials for blanking punches, controlled, head-to-head comparisons of their wear performance under identical conditions over extended stroke counts (up to 80,000) are scarce. Furthermore, existing physics-based wear models often require difficult-to-calibrate parameters, and data-driven machine learning (ML) approaches have not been systematically applied to predict volumetric punch wear as a function of stroke count and material properties in this specific context.

Methodology
The study employed a dual approach combining controlled experimentation with supervised machine learning:

Experimental Campaign:
- Materials: AISI D2 (1.5% C, 11.6% Cr, 0.7% Mo) and AISI D3 (2.1% C, 10.5% Cr, no Mo) punches were used to blank 1 mm thick ASTM A1011 low-carbon steel.
- Conditions: Trials were conducted on a 500-ton hydraulic press under dry-contact conditions (no lubricant) at 8 strokes/min. The punch-die clearance was fixed at 10% of sheet thickness (0.1 mm).
- Measurement: Punches were measured at intervals of 10,000 strokes (up to 80,000), with an additional early measurement at 5,000 strokes for D2. Cutting-edge radius was measured via optical microscopy, and volumetric wear ( $W_v$ ) was calculated using an analytical geometric model. Wear mechanisms were characterized via microscopy.
- Dataset: The initial dataset comprised 15 records (8 for D2, 7 for D3).
Machine Learning Framework:
- Data Augmentation: To address the small sample size, a Gaussian-noise augmentation strategy expanded the dataset from 15 to 150 records (10 synthetic replicates per original record) to prevent overfitting.
- Models: Three supervised regression models were developed and benchmarked:
  - Random Forest (RF): An ensemble of 200 decision trees.
  - Long Short-Term Memory (LSTM): A sequential deep learning network with two stacked layers (64 and 32 units) and dropout regularization.
  - Support Vector Regression (SVR): Using a Radial Basis Function (RBF) kernel.
- Features: Input variables included stroke count, hardness (HRC), ultimate tensile strength, and elastic modulus. The target variable was wear volume.
- Validation: Models were evaluated using 5-fold cross-validation and a held-out test set, with performance metrics including $R^2$ , RMSE, and MAE.

Key Results

Wear Performance: AISI D3 demonstrated superior wear resistance compared to AISI D2. At 80,000 strokes, the D3 punch exhibited a cutting-edge radius approximately 40% lower (93.97 µm vs. 156.7 µm) and a wear volume approximately 14.5% lower (8.30 × 10⁶ vs. 9.70 × 10⁶ µm³) than D2.
Wear Mechanisms: Microscopy revealed that AISI D2 suffered from combined adhesive and abrasive wear, characterized by Built-Up Edge (BUE) formation beyond 50,000 strokes. In contrast, AISI D3 maintained predominantly abrasive wear without significant material accumulation, preserving a cleaner cutting edge.
Convergence of Wear: The wear-volume ratio (D2/D3) was 4.77 at 10,000 strokes but narrowed significantly to 1.17 at 80,000 strokes. This convergence is attributed to BUE formation on D2, which initially masks wear volume before subsequent delamination accelerates apparent wear.
Model Performance: The Random Forest model achieved the highest predictive accuracy ( $R^2 = 0.987$ , RMSE = 142,300 µm³), outperforming LSTM ( $R^2 = 0.981$ ) and SVR ( $R^2 = 0.962$ ).
Feature Importance: Analysis identified stroke count as the dominant predictor of wear volume (58% importance), followed by hardness (22%), tensile strength (12%), and elastic modulus (8%).

Significance and Claims
The paper claims to provide a practical, generalizable framework for predictive tool-life planning and material selection in precision blanking. The study demonstrates that:

Material Selection: AISI D3 offers a quantifiable wear-resistance advantage for production runs below 50,000 strokes. However, for runs exceeding 60,000–80,000 strokes, the volumetric wear advantage diminishes, suggesting that selection criteria should shift to include burr formation and cut-edge surface quality, where D3 maintains a distinct advantage due to the absence of BUE.
Predictive Capability: The combined experimental and data-driven approach offers a computationally inexpensive route to forecast wear volume without requiring full mechanistic re-derivation for every new configuration. The Random Forest model is highlighted as particularly effective for interpolative accuracy within the tested stroke range.
Limitations and Future Work: The authors modestly note that the Random Forest model's reliability decreases when extrapolating beyond the 80,000-stroke training limit. They suggest that future work should incorporate additional process variables (clearance, lubrication), other tool-steel grades, and online sensor data (acoustic emission, force-displacement) to enable real-time monitoring and hybrid physics-informed modeling. The study does not claim to solve all tool-life prediction challenges but offers a validated step toward condition-based maintenance in blanking operations.

More like this