The core challenge the researchers set out to solve is uncertainty. In an idealized computer model, bearing stiffness and damping are tidy numbers pulled from a manufacturer’s catalog. In the real world, they are anything but. Manufacturing tolerances, assembly errors, thermal expansion mismatches between the steel shaft and its housing, and speed-dependent centrifugal forces all conspire to shift these parameters by roughly 5 to 12 percent from their nominal values. Traditional deterministic designs may look excellent on paper but can fluctuate wildly or fail outright when reality deviates from the assumed values. Worse, most prior studies have relied on theoretical catalog data rather than measurements, ignoring what engineers call epistemic uncertainty, the gap between the model and the physical machine.

Wu and Ma’s first move was to measure rather than assume. They built a high-speed motorized spindle test rig spanning an operational range from 1,000 to 24,000 rpm, discretized into 50 speed points. At each point, the spindle ran under steady-state thermal conditions while high-precision eddy current sensors sampled vibration at 10 kilohertz. Repeating the identification procedure 20 times per speed produced a dataset of 1,000 identified parameter vectors. To extract bearing stiffness and damping coefficients from these measured frequency responses, the team employed the Multi-Innovation Stochastic Gradient, or MISG, algorithm, a recursive estimation technique that accelerates convergence by stacking innovation information from several consecutive time steps rather than using a single sample at a time.

The numerical verification of this algorithm is striking in its own right. In synthetic tests where the true parameter values were known, MISG reduced the number of convergence iterations from 606 to just 75, an 8.1-fold speedup and an 87.6 percent reduction compared with the standard stochastic gradient method, while also lowering the final normalized root-mean-square error by 15 percent. Although each iteration costs about 18 percent more computation, the faster convergence cuts total computation time by 28 percent. A sensitivity study further confirmed that algorithmic identification error contributes less than 3 percent of the total output variance, meaning the scatter in the identified parameters reflects genuine physical variability, not artifacts of the estimation procedure.

With the empirical distributions in hand, the team validated them statistically before feeding them into the optimizer. The identified stiffness parameters showed coefficients of variation between roughly 6.0 and 6.4 percent, and Shapiro-Wilk tests confirmed that a multivariate normal distribution was a physically permissible model, with quantile-quantile plots yielding coefficients of determination above 0.99. Bearing damping was not treated as an independent random variable; instead, it was coupled proportionally to the identified stiffness through a constant calibrated by modal hammer testing, keeping the probabilistic model lean while preserving fidelity. The identified stiffness values deviated only about 1.5 to 3.0 percent from the numerical baseline, with the small residual attributed to installation-specific effects such as thermal expansion mismatch and bearing-seat tolerances, exactly the factors catalog values ignore.

The optimization phase then framed spindle design as a bi-objective problem: minimize both the mean peak vibration amplitude at the tool tip and its standard deviation, subject to a probabilistic constraint that the probability of vibration exceeding a critical threshold of 195 micrometers stays below a target level. That threshold was calibrated from surface roughness requirements in precision milling, where larger displacements degrade finish beyond aerospace-grade limits. Rather than relying on first-order or second-order reliability methods, which linearize the problem and can incur errors of 20 to 30 percent in nonlinear bearing-rotor dynamics, the researchers propagated uncertainty through direct Monte Carlo simulation with 1,000 samples per candidate design. The Non-dominated Sorting Genetic Algorithm II then evolved populations of designs, with each run producing a Pareto front of optimal trade-offs.

Crucially, the team ran the optimization three times, for target failure probabilities of 1, 5, and 10 percent, generating a family of Pareto fronts that reveals the marginal cost of safety. The results expose sharp diminishing returns. Tightening the constraint from 10 to 5 percent delivered a high marginal benefit ratio of 0.55 micrometers of variance reduction per percentage point of reliability, but squeezing from 5 to 1 percent dropped that ratio to 0.39 and demanded a 7.0 percent increase in bearing preload, pushing it to its physical limit of 3,000 newtons, for only a marginal robustness gain. Such aggressive preload accelerates rolling-element wear and risks friction-induced thermal problems. Weighing these trade-offs against ISO 20816-3 vibration standards, the researchers identified the 5 percent configuration as the optimal engineering equilibrium.

The optimized design itself is physically revealing. The framework selected a front-bearing preload of about 2,804 newtons and shortened the bearing support span from roughly 750 to 537 millimeters, a geometric change that raises the first natural frequency and pushes the third critical speed from about 22,000 rpm to 25,800 rpm, safely beyond the 24,000 rpm maximum operating speed. Benchmarking against conventional approaches underscored the value of the joint formulation: a pure robust design optimization achieved low variability but left the failure probability at an unacceptable 21.07 percent, while a pure reliability-based design hit a low failure probability but suffered an inflated response standard deviation of 11.80 micrometers. The integrated framework achieved failure probability of 0.35 percent while keeping variability at 9.69 micrometers and mean amplitude at 16.44 micrometers.

Physical validation sealed the case. Twenty independent replication runs at the rated speed of 24,000 rpm showed the optimized spindle reducing mean peak vibration by 16.84 percent, from 19.77 to 16.44 micrometers, and the response standard deviation by 29.78 percent, from 13.80 to 9.69 micrometers, with all improvements statistically significant in paired-sample t-tests. Across the full 1,000 to 24,000 rpm spectrum, the optimized spindle maintained stable, constrained responses, eliminating the severe resonant amplification that plagued the baseline design near its third critical speed. The variance reduction of 50.70 percent in experimental terms also surpasses the roughly 22 percent typically reported in simulation-based robust optimization studies.

Beyond the immediate numbers, the study signals a broader shift in how rotating machinery can be engineered. By grounding probabilistic models in measured reality rather than catalog assumptions, and by mapping the full trade-off surface between performance, robustness, and reliability instead of fixing a single safety target, the framework gives engineers a quantitative language for deciding how much safety margin a given application truly needs. The authors note that the approach currently assumes linear bearing behavior within the Hertzian elastic regime and depends on sufficient experimental data, pointing toward amplitude-dependent identification schemes and non-probabilistic interval methods as future extensions for heavily loaded or multi-shaft systems. For now, the message is clear: the path to quieter, more dependable high-speed spindles runs directly through the test rig, with every ounce of uncertainty measured, quantified, and designed against.

Subject of Research: Reliability-based robust optimization of high-speed motorized spindle-bearing systems using experimentally identified bearing parameter uncertainties

Article Title: Experimental identification and robust optimization of spindle–bearing systems with reliability constraints

Article References: Wu, H., & Ma, J. (2026). Experimental identification and robust optimization of spindle–bearing systems with reliability constraints. Mechanical Sciences, 17(2), 839-855. https://doi.org/10.5194/ms-17-839-2026

Image Credits: AI Generated

DOI: 10.5194/ms-17-839-2026

Keywords: motorized spindle, bearing stiffness, vibration control, robust optimization, reliability-based design, uncertainty quantification, Monte Carlo simulation, NSGA-II, multi-innovation stochastic gradient, CNC machining, Pareto optimization, rotor dynamics

The study, published in Results in Engineering by Kazuki Kaneko and Zongze Li, addresses a fundamental blind spot in digital manufacturing. Computer-aided manufacturing (CAM) software can simulate geometry and generate tool paths from a CAD model, but it cannot simulate what happens inside the machine’s controller. The G-code is filtered by the CNC interpolator, feedforward controllers and feedback control loops before it ever reaches the servomotors, and these proprietary algorithms are closely guarded black boxes maintained by machine tool builders. Physical modeling of this chain is extraordinarily difficult precisely because the control algorithms are secret and the dynamics are deeply nonlinear. The result is that manufacturers often discover machining errors only after cutting — or resort to time-consuming air runs in which the machine executes the program without a workpiece just to observe its behavior.

The authors reframed the problem as one of sequence translation. From a data-processing perspective, G-code execution is a stream of discrete commands fed row by row into a black-box system that emits corresponding sequences of tool-path and cycle time errors. Predicting those errors for a given program is therefore fundamentally a sequence-to-sequence problem — exactly the kind of task at which Transformers excel. Unlike recurrent networks or long short-term memory (LSTM) models, which process sequences step by step, a Transformer uses multi-head self-attention to derive global relational weights across all positions simultaneously. This allows it to capture long-range temporal dependencies, such as how a sharp corner several blocks ahead causes the servo system to decelerate and how a large acceleration from a previous segment continues to influence subsequent commands as the transient response gradually decays.

The researchers drew a striking analogy between the Transformer’s internal machinery and the physics of a CNC controller. In their framing, the query represents the motion state of the current G-code command, while the keys and values carry error features of the surrounding tool path; scaled dot-product attention then assigns weights to nearby and distant blocks, mirroring the feedforward and feedback structure of a real servo loop. Position-wise feedforward networks introduce the nonlinear modeling capacity needed to emulate local servo behaviors such as PID control and friction, while residual connections — which pass inputs directly forward — resemble a closed-loop feedback system that focuses on predicting small dynamic deviations rather than reconstructing the entire tool path. Layer normalization, meanwhile, stabilizes the wildly different physical scales of position increments and feed rates, functioning much like a gain stabilizer in a servo drive.

Training such a model requires data that spans the machine’s full dynamic repertoire, so the team avoided reliance on specific workpiece geometries, which would limit generalization. Instead, they semi-randomly generated ten sets of training G-code combining local blending segments, where the path passes directly through control points, and global blending segments, where B-spline interpolation smooths the trajectory at constant feed rate. Displacements and feed rates were drawn randomly from defined ranges, producing tool paths with varied travel ranges and frequent direction changes. Each program was then executed on a real three-axis ROBODRILL machine equipped with a Fanuc Series 31i controller, and actual tool-path data were captured at a sampling frequency of 1 kHz using Fanuc’s SERVO Guide monitoring software. The training targets were thus experimentally measured servo responses rather than virtual simulations, grounding the digital twin in physical reality.

The model itself is an encoder-only Transformer with a model dimension of 512, eight layers and four attention heads, trained for 2000 epochs on a workstation with an RTX 4090 GPU. Input sequences consisted of nominal positional increments along the three feed axes plus the commanded feed rate, segmented into windows of 32 control points using a causal sliding window. The output comprised the three-dimensional tool-path error and the cumulative cycle time error at each point. Notably, the authors found that longer input sequences converged more slowly without improving final accuracy — a result that aligns neatly with servo physics, since the transient influence of any single command naturally fades over time. When compared against a Bi-LSTM model of equal dimensions trained on identical data, the Transformer converged in roughly 500 epochs, three times faster than its recurrent rival.

Validation was carried out on six G-code programs derived from two physical specimens: a 25-millimeter-radius hemispherical part machined along morphing spiral and bidirectional paths, and a complex artifact geometry, each run at nominal feed rates of 1000 and 3000 millimeters per minute. The results were decisive. Both models avoided large systematic bias, with mean prediction deviations generally within ±1 micrometer or millisecond, but the Transformer’s uncertainty bands were roughly 1.5 to 2 times tighter than the Bi-LSTM’s across all error components — and in some conditions nearly four times tighter. Its error distributions clustered markedly more sharply around zero, and while the Bi-LSTM’s cycle time root-mean-square error ballooned to about 8 milliseconds on the most demanding program against a mean error of 2.5 milliseconds, the Transformer suppressed such outliers, keeping RMSE close to MAE.

The comparison also revealed how differently the two architectures degrade under stress. Higher feed rates inherently reduce the accuracy of the controller’s feedforward and feedback loops, and the Bi-LSTM’s errors doubled or even tripled when the feed rate rose from 1000 to 3000 millimeters per minute. The Transformer, by contrast, showed far more graceful degradation, keeping mean errors below 2 micrometers or milliseconds even on the most challenging high-speed program. The authors attribute the residual errors that remain — small but nonzero — to highly random mechanical nonlinearities such as micro-friction variations and backlash during axis reversals, along with the quantization limits of encoder feedback at the 1 kHz sampling rate.

Perhaps the most industrially compelling result is speed. Predicting errors for an entire validation program took the Transformer less than 3.5 seconds, regardless of program complexity, whereas air-running the same programs on the machine took anywhere from roughly 38 to 202 seconds. Because prediction time depends only on G-code length, the trained model can instantly screen massive production programs offline, eliminating the need for preliminary air runs and sharpening both production scheduling and machining efficiency. The approach does not generalize across machines — each machine tool has unique controller and mechanical characteristics — but the authors emphasize that adapting to a new machine requires only executing the training G-codes, acquiring a fresh dataset and retraining, after which the model serves as a customized digital surrogate for that specific machine.

The work points toward a future in which every machine tool carries its own learned digital twin, capable of vetting programs before metal is ever cut. The team’s next steps include extending the model to account for cutting force-induced deflections of tool and workpiece, which superimpose on servo tracking errors in real machining; probing how specific G-code features such as control point density affect prediction performance; scaling the sequence-to-sequence methodology to five-axis simultaneous machining; and developing real-time servo signal acquisition for online error prediction during actual cutting. They also plan to build open-architecture experimental CNC systems, since proprietary commercial controllers block the integration of theoretical control equations needed for hybrid physical-digital modeling. If those ambitions pan out, the humble G-code file may one day arrive at the shop floor already annotated with a complete forecast of its own imperfections.

Subject of Research: Transformer-based deep learning prediction of tool-path and cycle time errors in CNC machining from G-code

Article Title: Prediction of tool-path and cycle time errors of CNC machining using a transformer-based deep learning model

Article References: Kaneko, K., & Li, Z. (2026). Prediction of tool-path and cycle time errors of CNC machining using a transformer-based deep learning model. Results in Engineering, 32, Article 112872. https://doi.org/10.1016/j.rineng.2026.112872

Image Credits: AI Generated

DOI: 10.1016/j.rineng.2026.112872

Keywords: CNC machining, Transformer model, deep learning, digital twin, tool-path error, cycle time prediction, G-code, servo dynamics, Bi-LSTM comparison, precision manufacturing, sequence-to-sequence learning, machine tools