⚡ electrical engineering

Deep Learning-Driven Multi-Objective Optimization of Surface Finish and Tool Wear in CNC Dry Turning

This paper presents a deep learning-driven multi-objective optimization framework that outperforms classical power-law models in predicting surface finish and tool wear for CNC dry turning, successfully identifying a superior Pareto front and robust operating parameters through NSGA-III and TOPSIS analysis.

Original authors: Shuma Fadhili, Mathias Sebastian Halinga, Haryson Johanes Nyobuya

Published 2026-08-19

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Shuma Fadhili, Mathias Sebastian Halinga, Haryson Johanes Nyobuya

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 world of manufacturing, the machine shop is a place of constant negotiation. A lathe, a machine that spins metal to shape it, must be guided by a human operator who faces a difficult choice every time they start a cut. They want the finished part to be perfectly smooth, free of the tiny scratches that can cause leaks or fatigue in engines and tools. Yet, the very settings that create that smoothness often chew through the cutting tool much faster, forcing the factory to buy expensive replacements. This trade-off between a fine surface and a long-lasting tool is the central problem of turning metal. For decades, operators have relied on old handbooks and conservative guesses to find a middle ground, hoping to avoid wasting money on ruined parts or broken tools. But the physics of cutting metal is complex and nonlinear; the relationship between how fast the machine spins, how fast it feeds the metal, and how deep it cuts does not follow simple, straight lines. When the rules are this complicated, old guesses often fail, leading to parts that are either too rough or tools that wear out too soon.

A team of researchers at the University of Dar es Salaam decided to solve this problem not with more guesswork, but with a modern approach that lets the data speak for itself. They set out to build a digital guide for a specific type of heavy-duty lathe used in Tanzanian factories, one that could tell an operator exactly how to set the machine to get the best possible result. Instead of relying on the simple formulas found in textbooks, they used a type of artificial intelligence known as a deep neural network. Think of this technology as a highly flexible brain that can learn complex patterns from a set of examples, rather than a rigid calculator that only follows a single, fixed rule. The researchers wanted to see if this flexible brain could find a better path through the trade-off between surface smoothness and tool wear than the traditional methods ever could.

To teach their system, the team ran a carefully planned series of twenty-seven experiments on a production-grade CNC lathe. They used bars of medium-carbon steel, a material common in the region, and a single type of coated carbide cutting tool. They varied three main settings for each run: the speed at which the metal spun, the rate at which the tool moved along the metal, and the depth of the cut. After each run, they measured two things with high precision: the roughness of the surface left behind and the amount of wear on the cutting edge of the tool. This small but complete set of data became the foundation for their study. They then split their work into two paths. On one path, they fitted the data to the traditional, simple mathematical formulas that factories have used for years. On the other path, they trained their deep neural network to learn the patterns hidden within the same twenty-seven numbers.

The results revealed a clear difference in how well each method understood the machine. The traditional formulas, while decent at predicting surface roughness, struggled significantly when it came to tool wear. The researchers found that the wear on the tool was driven by a complex mix of interactions between the settings, a structure that the simple formulas could not capture. Because the formulas missed these subtle connections, they made a serious error in the most critical area: they told the operator that the smoothest possible finish was much rougher than it actually was. Specifically, the old model predicted a minimum roughness value that was forty percent higher than what the machine actually achieved in the lab. This was not just a small mistake; it meant the traditional guide was hiding the best possible outcomes from the operator, effectively telling them that a perfect finish was impossible when it was actually within reach.

The deep neural network, by contrast, saw the full picture. It learned the complex interactions that the simple formulas missed and predicted the tool wear with far greater accuracy. When the researchers used this smarter model to search for the best possible settings, the results changed dramatically. The new guide showed a path to a finish that was twenty-eight percent finer than the old guide ever suggested. It uncovered a "sweet spot" where the machine could produce a surface finish of 2.391 micrometers, a level of smoothness that the traditional model had completely overlooked. The study proved that by using a flexible learning model instead of a rigid formula, the researchers could expose a much wider range of high-quality options that were previously invisible to the factory floor.

With the better map in hand, the team then had to decide which of these many options was the best recommendation for a real-world operator. They faced a classic dilemma: the settings that gave the absolute smoothest finish wore out the tool faster, while the settings that saved the tool left the surface slightly rougher. To solve this, they used a decision-making method that finds the "knee" of the curve, the point where you get the most benefit for the least extra cost. They found that for this specific machine and material, the feed rate and the depth of the cut should always be set to the same two specific values, regardless of the goal. The only setting that needed to change was the spindle speed. This simplified the operator's job immensely. Instead of juggling three variables, they only needed to turn one dial.

The researchers also tested how sensitive this recommendation was to the operator's priorities. They asked: what if the factory cares more about speed than finish, or vice versa? They found that the recommendation was surprisingly robust. As long as the operator did not prioritize the surface finish above fifty-five percent of the total importance, the best setting remained the same: a slower speed that saved the tool. Only when the finish became the absolute top priority did the recommendation jump to the fastest speed. This stability means that a factory does not need to constantly recalculate settings for every slight change in preference; they can rely on a single, solid recommendation for most production needs.

The study concludes that the old way of relying on simple formulas is leaving money and quality on the table. By using a deep learning model trained on real-world data from a production machine, the researchers provided a tool that is not only more accurate but also more honest about what the machine can do. They showed that the bias in the old models was not just a theoretical error but a practical one that distorted the menu of choices available to the operator. The new approach offers a clear, data-driven path forward, turning a complex, confusing trade-off into a simple, reliable instruction that any operator can follow to get better parts and save money on tools.

Technical Summary: Deep Learning-Driven Multi-Objective Optimization of Surface Finish and Tool Wear in CNC Dry Turning

Problem Statement
In CNC turning, achieving a fine surface finish ( $R_a$ ) without sacrificing tool life (flank wear, $V_B$ ) presents a critical trade-off. While classical parameter selection often relies on conservative handbook values or low-order empirical models, these approaches struggle with the nonlinear interactions between cutting parameters (spindle speed $N$ , feed rate $f$ , depth of cut $d$ ) and process physics. In resource-constrained manufacturing environments, such as those in Tanzania, the cost of tooling and the scarcity of foreign currency for replacements make suboptimal parameter selection economically significant. Existing literature often separates predictive modeling from optimization or relies on laboratory-grade machines, leaving a gap in validated, production-grade decision tools that account for surrogate model bias in multi-objective optimization.

Methodology
The authors propose a six-stage, fully data-driven framework validated on a CK6166 production CNC lathe machining AISI 1045 medium-carbon steel under dry conditions.

Experimental Design: A randomized $3^3$ full factorial design (27 runs) was executed, varying $N$ (200–600 rpm), $f$ (0.2–1.0 mm/rev), and $d$ (0.5–2.0 mm). Responses measured were surface roughness ( $R_a$ ) and flank wear ( $V_B$ ).
Statistical Screening: Analysis of Variance (ANOVA) was used to identify significant main effects and interactions, establishing the statistical structure of the objective conflict.
Modeling: Two model families were developed in parallel:
- Classical Baseline: Multiplicative power-law regressions fitted via log-linearization.
- Deep Learning Surrogates: Compact Deep Neural Networks (DNNs) with a 3–16–16–1 architecture, trained in log-space using Adam optimization and L2 regularization to handle the small dataset (27 observations).
Benchmarking: Both models were evaluated using Leave-One-Out Cross-Validation (LOOCV) to assess generalization rather than in-sample fit.
Optimization: Single-objective bounds were established using Differential Evolution (DE). Multi-objective optimization was performed using the reference-point-based NSGA-III algorithm to trace the Pareto front.
Decision Analysis: The Pareto front was converted into operator-ready recommendations using the geometric knee-point method and TOPSIS (Technique for Order of Preference by Similarity to Ideal Solution) ranking, including a weight-sensitivity analysis.

Key Results

Statistical Drivers: ANOVA confirmed that surface roughness is overwhelmingly dominated by feed rate (97.19% contribution), while tool wear is governed primarily by depth of cut (67.21%) and spindle speed (17.75%). Crucially, tool wear exhibited significant interaction effects ( $N \times d$ and $f \times d$ ) accounting for over 7% of variance, which power-law models cannot capture.
Model Performance: The DNN surrogates outperformed power-law models in LOOCV.
- For $R_a$ , DNN reduced Mean Absolute Percentage Error (MAPE) from 9.79% to 7.65%.
- For $V_B$ , DNN reduced MAPE from 6.68% to 1.01% (a six-fold improvement).
Impact on Optimization: The choice of surrogate significantly altered the Pareto front. The power-law model overestimated the minimum achievable roughness by 40% (predicting 3.34 $\mu$ m vs. the experimental 2.391 $\mu$ m) and compressed the apparent trade-off range by 28%. The DNN-driven front extended deeper into the fine-finish region, recovering the experimentally observed optimum.
Optimal Parameters: Across the Pareto front, feed rate and depth of cut collapsed to single robust values ( $f=0.2$ $f = 0.2$ mm/rev, $d=0.5$ $d = 0.5$ mm). The entire trade-off was governed solely by spindle speed.
- Knee Point Recommendation: $N=385$ rpm, yielding $R_a \approx 3.70$ $\mu$ m and $V_B \approx 0.011$ mm.
- TOPSIS Sensitivity: Under equal weights, TOPSIS favored the wear-economical extreme ( $N=200$ rpm). However, the recommendation remained stable at $N=200$ rpm for all roughness weights up to 0.55, indicating robustness in decision-making.

Significance and Claims
The paper claims three primary contributions:

Validated Pipeline: It delivers a complete, experimentally validated deep-learning optimization pipeline specifically for production-grade turning, bridging the gap between predictive modeling and actionable parameter recommendations.
Quantification of Bias: It provides the first quantified evidence for this process that surrogate model bias directly distorts the Pareto front presented to decision-makers. The study demonstrates that a biased parametric model does not just mispredict values but "misprices" every trade-off, potentially leading to suboptimal operational choices.
Practical Deployment: The framework is computationally lightweight, requiring no GPU or proprietary software, and reduces the complex multi-variable problem to a single spindle-speed dial for operators, making it suitable for deployment in resource-constrained manufacturing environments.

The authors maintain modesty regarding limitations, noting that the models are valid strictly within the tested factor box, apply to dry cutting of AISI 1045, and that the lightest wear measurements approach the resolution limits of optical microscopy. They emphasize that the findings are specific to the experimental conditions and that dedicated confirmation experiments at the knee point are recommended before contractual adoption.

More like this