⚡ electrical engineering

Physics-guided machine learning for trochoidal milling: force and power prediction and limits of cross-diameter generalisation

This paper demonstrates that a compact, physics-informed machine learning model incorporating maximum uncut chip thickness and a simplified severity descriptor effectively predicts mean cutting forces and spindle power in trochoidal milling of AA7075-T651, offering a robust and interpretable alternative to complex models for preliminary process screening despite limitations in capturing transient peak forces.

Original authors: Gorka Urbikain Pelayo, Daniel Olvera-Trejo, Luis Norberto López de Lacalle, Alex Elías-Zuñiga

Published 2026-10-16

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Gorka Urbikain Pelayo, Daniel Olvera-Trejo, Luis Norberto López de Lacalle, Alex Elías-Zuñiga

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 high-stakes world of aerospace manufacturing, engineers face a constant tug-of-war between speed and safety. They need to carve massive, intricate pockets out of tough metal frames as quickly as possible, yet they must avoid breaking the delicate tools or the thin walls of the aircraft parts. To solve this, machinists often use a strategy called trochoidal milling. Instead of plunging a spinning cutter straight into the metal like a drill, the tool follows a looping, circular path while moving forward. This motion allows the tool to take small, manageable bites of material, keeping the load on the cutter low and preventing it from overheating or snapping. However, because the tool is constantly entering and leaving the metal in a complex dance of curves, predicting exactly how much force it will exert or how much power the machine will need is incredibly difficult. If the prediction is wrong, the tool breaks, the part is ruined, or the machine wastes energy.

A team of researchers set out to bring clarity to this chaotic process by combining the laws of physics with modern data analysis. They focused on a specific, high-strength aluminum alloy used in aircraft, testing it with three different sizes of cutting tools. Their goal was not just to measure what happened, but to build a smart model that could predict the outcome of future cuts without needing to run a physical test every time. They gathered data from seventy-two separate cutting experiments, carefully tracking the forces pushing against the tool and the electricity consumed by the machine's motor. By feeding this data into a computer program designed to learn from patterns, they tested different ways of describing the cutting conditions to see which description allowed the computer to make the most accurate guesses.

The researchers discovered that the size of the tool matters more than one might expect, but not in a simple way. When they used the largest tool, measuring twelve millimeters across, the forces pushing sideways on the tool were significantly higher than with the smaller eight-millimeter tool. Surprisingly, the largest tool actually required less power from the machine motor on average. This counterintuitive result happened because the larger tool was programmed to follow a wider, gentler loop, which smoothed out the energy demand even though the physical force on the metal was greater. The study also revealed that the speed at which the tool fed into the metal had a clear effect on power consumption but surprisingly little effect on the sideways forces, suggesting that operators can adjust the feed rate to manage energy use without worrying about overloading the tool.

To make sense of these complex interactions, the team tried three different approaches to feeding information into their computer model. The first approach used only the most basic settings, like tool size and feed rate. The second approach added two specific physical concepts: the maximum thickness of the tiny chip of metal being cut and a simplified measure of how intense the cutting process was overall. The third approach tried to add even more raw data about the depth and width of the cut. The results showed that the middle approach was the winner. By including the chip thickness and the intensity measure, the model became significantly better at predicting both the average force on the tool and the average power needed. Adding the extra raw data from the third approach did not help; it only introduced confusion because that information was already hidden inside the other variables.

The study also highlighted a crucial difference between predicting average behavior and predicting sudden spikes. The computer model was quite good at guessing the average force and the average power, which are stable, steady-state values. However, it struggled to predict the maximum peak forces that occur for a split second during the cut. This is particularly true for the force pushing straight down on the tool, which the model failed to predict better than simply guessing the average. The researchers concluded that these sudden, sharp peaks are caused by fleeting, local events—like a tiny vibration or a momentary snag in the metal—that are too fast and too specific to be captured by general descriptions of the cut.

Perhaps the most significant finding was a limitation in how far these predictions can travel. The model worked well when it was asked to predict results for tool sizes it had already seen in the training data. However, when the researchers tried to use the model to predict results for a tool size it had never encountered, the predictions collapsed. The model could not simply "learn" the rules of the game and apply them to a new tool diameter. This suggests that while the computer can help plan cuts for known tool sizes, it cannot yet replace the need for physical testing when introducing a completely new tool size. The study confirms that for the specific aluminum alloy and tool sizes tested, a compact set of physical descriptors can help engineers screen for efficient cutting conditions, but it also draws a clear line where the data-driven approach ends and the need for real-world experimentation begins.

Technical Summary: Physics-guided Machine Learning for Trochoidal Milling

Problem Statement
Trochoidal milling has emerged as a critical strategy for the high-efficiency machining of large pockets and thin-walled features in lightweight aerospace structures, particularly for difficult-to-cut alloys like AA7075-T651. However, predicting cutting forces and spindle power in this process is challenging due to the continuously varying cutter-workpiece engagement and uncut chip thickness along the toolpath. Traditional mechanistic models require complex, real-time updates of engagement geometry, which limits their application across varying toolpaths and tool diameters. Furthermore, existing data-driven approaches often rely on single tool diameters or purely black-box models that lack physical interpretability. There is a specific gap in systematically comparing input representations across multiple tool geometries to determine which physics-informed descriptors best support predictive modeling for preliminary process screening.

Methodology
The study employed an experimental and data-driven approach involving 72 machining tests on AA7075-T651 aluminium alloy using three end-mill diameters (8, 10, and 12 mm).

Experimental Design: The campaign varied tool diameter, feed per tooth ( $f_z$ ), radial engagement ( $a_e$ ), axial depth of cut ( $a_p$ ), and orbital diameter ( $D_{orb}$ ). To ensure comparability across diameters, parameters were normalized using dimensionless descriptors: normalised axial engagement ( $\lambda$ ), normalised radial engagement ( $\mu$ ), and orbital-diameter ratio ( $\delta$ ).
Data Acquisition: Cutting forces (in-plane $F_{xy}$ and axial $F_z$ ) were measured synchronously with a piezoelectric dynamometer, while spindle power was logged via the machine monitoring system.
Feature Extraction: Five robust output indicators were extracted: mean in-plane force ( $F_{xy,mean}$ ), maximum in-plane force ( $F_{xy,max}$ ), robust peak-to-peak in-plane force ( $F_{xy,ptp}$ ), maximum axial force ( $F_{z,max}$ ), and mean spindle power ( $P_{mean}$ ).
Modeling Framework: Three physics-informed input structures were constructed and compared using Least Squares Boosted Regression Trees (LSBoost) with 5-fold cross-validation:
1. Minimal (5 variables): $D, \lambda, \mu, \delta, f_z$ .
2. Hybrid (5+2 variables): Minimal set plus maximum uncut chip thickness ( $h_{max}$ ) and a simplified severity descriptor ( $MRR^* = a_e \cdot a_p \cdot f_z$ ).
3. Enriched (5+4 variables): Hybrid set plus raw engagement variables ( $a_e, a_p$ ).
Validation: Models were evaluated using Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and the coefficient of determination ( $R^2$ ). A specific "leave-one-diameter-out" protocol was used to test cross-diameter generalization.

Key Results

Model Performance: The Hybrid (5+2) model provided the best overall compromise between predictive accuracy and input redundancy. It improved the $R^2$ for mean in-plane force ( $F_{xy,mean}$ ) from 0.241 (Minimal) to 0.364 and for mean spindle power ( $P_{mean}$ ) from 0.374 to 0.473. The Enriched (5+4) model performed similarly to or worse than the hybrid model for most targets, indicating that the additional raw variables mainly reintroduced redundancy rather than providing new information.
Predictability of Targets: Average, system-level responses were significantly more predictable than peak-based ones. $P_{mean}$ was the best-predicted target ( $R^2 = 0.473$ ), followed by $F_{xy,mean}$ . $F_{xy,ptp}$ was moderately predicted ( $R^2 = 0.263$ ), while $F_{xy,max}$ was weakly predicted ( $R^2 = 0.069$ ). Notably, all three model structures yielded negative $R^2$ values for maximum axial force ( $F_{z,max}$ ), indicating that test-level descriptors cannot capture the transient, local phenomena (e.g., vibration, runout) governing axial peaks.
Algorithm Comparison: Non-linear tree ensembles (LSBoost) outperformed multiple linear regression across all targets, confirming the non-linear nature of the process. However, the performance gap between LSBoost and bagged trees was small, suggesting that the primary benefit came from using a non-linear tree-based representation rather than the specific boosting algorithm.
Cross-Diameter Generalization: The models failed to generalize to unseen tool diameters. When trained on two diameters and tested on the third, the mean $R^2$ dropped to approximately -2.6. Removing the raw tool diameter from the inputs did not improve this, suggesting that normalized descriptors alone cannot fully capture diameter-dependent changes in stiffness and dynamic response.

Significance and Claims
The paper claims that compact, physically motivated feature sets can effectively support the preliminary screening of trochoidal milling conditions without resorting to larger, less interpretable models. Specifically, the study demonstrates that:

Hybrid descriptors are superior: Adding physically derived severity indicators ( $h_{max}$ and $MRR^*$ ) to a minimal geometric set improves prediction of average forces and power, whereas adding raw engagement variables creates redundancy.
Limitations of global descriptors: The inability to predict axial force peaks ( $F_{z,max}$ ) serves as a diagnostic result, highlighting that global, test-level inputs are insufficient for capturing local, transient mechanical events.
Scope of applicability: The proposed models are suitable for interpolation within the range of tested tool diameters but are not currently generalizable to entirely new tool sizes. The authors conclude that while the hybrid model offers a favorable balance of performance and complexity for process planning, its application is restricted to conditions represented in the training data.

The study does not claim to solve the problem of cross-diameter generalization or peak-force prediction but rather defines the boundaries of current data-driven approaches in this domain, emphasizing the need for future work involving time-resolved engagement descriptors and additional experimental variables (runout, vibration) to address these limitations.

More like this