Enhancing Experimental Efficiency in Materials Design: A Comparative Study of Taguchi and Machine Learning Methods

Antov Selwinston; P Akshay Kumar; Pavan Taduvai; Rohit Batra; Shreya Bairi; Shyam Prabhu

REVIEW 4 major objections 5 minor 12 references

Enhancing Experimental Efficiency in Materials Design: A Comparative Study of Taguchi and Machine Learning Methods

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read An active-learning Gaussian process model beats the Taguchi L25 method on weld bead prediction using fewer experiments.

desk verdict Real WAAM data and a sensible comparison, but the main claim is only as good as the unspecified test-set exclusion; fix that before trusting the numbers. read the letter →

arxiv 2506.03910 v1 pith:IRWYG26U submitted 2025-06-04 cs.LG

Shyam Prabhu , P Akshay Kumar , Antov Selwinston , Pavan Taduvai , Shreya Bairi , Rohit Batra This is my paper

classification cs.LG

keywords TaguchimethodGaussianprocessregressionactivelearningLatinhypercubesamplingwirearcadditivemanufacturingbeadgeometrydesignofexperimentsmaterials

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper claims that an active-learning Gaussian process regression (GPR) model can predict weld bead geometry in wire-arc additive manufacturing more accurately, and with fewer physical experiments, than the standard Taguchi design-of-experiments method. The authors compare a fixed L25 orthogonal-array Taguchi design, which consumes 25 experiments, against a GPR that starts from five Latin-hypercube samples and adds one experiment at a time where its predictive uncertainty is highest. On 15 test combinations, the GPR trained on 20 points reaches an $\mathrm{RMSE}$ of 1.00 and an $R^2$ of 0.89, beating the Taguchi model's $\mathrm{RMSE}$ of 1.33 and $R^2$ of 0.80. The learning curve shows the GPR matching Taguchi's accuracy after about seven experiments and surpassing it after 17. This matters because experimental cost often dominates materials design, so a method that is both more accurate and more sample-efficient would let practitioners map a process window with fewer builds.

What carries the argument

The load-bearing machinery is a Gaussian process regression surrogate with a radial-basis-function kernel, a smooth similarity measure between process settings, started with five Latin-hypercube samples and refined by an uncertainty-based active-learning loop: after each fit, the next experiment is chosen at the input point where the model's predictive variance is largest. This is compared against the Taguchi L25 orthogonal array, a fixed 25-run design whose predictions come from an additive main-effects formula in the process variables. The GPR's uncertainty estimate is what lets it concentrate experiments in informative regions, while the Taguchi design cannot revise its sampling plan in response to measured data.

What would settle it

Rerun the comparison with the 15 test cases chosen by a fixed random seed or by stratified sampling across all five levels of each factor, keeping the same LHS initialization and active-learning loop; if the 20-point GPR then fails to beat the L25 Taguchi model on $\mathrm{RMSE}$ or $R^2$, the claimed superiority does not generalize beyond the paper's particular test selection.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that, for this three-factor, five-level WAAM bead-geometry problem, a Gaussian process with uncertainty-based active learning dominates the Taguchi L25 method on both accuracy and data efficiency. The final GPR model (length scale 2.68, noise level 1.70, radial-basis-function kernel) used only 20 experimental points and scored an $\mathrm{RMSE}$ of 1.00 and an $R^2$ of 0.89 on 15 held-out test cases, while the Taguchi model built from 25 points scored 1.33 and 0.80. Taguchi's predictions showed systematic underprediction, while GPR errors were centered near zero, evidence that the Taguchi main-effects formula misses nonlinear interactions among current, weld speed, and contact-tip-to-work distance. The learning curve demonstrates that GPR matches Taguchi's accuracy with roughly seven experiments and overtakes it with 17.

Load-bearing premise

The load-bearing assumption is that the 15 test combinations are a representative, unbiased sample of the 125-point design space; the paper says only that they were 'selected as test cases' from that space, with no stated random seed, stratification, or exclusion rule.

Editorial extensions

If this is right

A practitioner could replace the standard 25-run Taguchi plan with a 17-experiment active-learning loop and expect better or equal predictive accuracy for this WAAM bead-geometry task.
The same GPR-plus-active-learning construction should transfer to other continuous multi-factor materials processing problems, because nothing in the method relies on welding-specific physics.
Because GPR provides a predictive distribution, the framework can be extended from prediction to optimization, for example by using expected-improvement acquisition to search for process parameters that give a target bead geometry.
The systematic underprediction error of the Taguchi baseline indicates that linear main-effects designs are insufficient when process variables interact nonlinearly.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

A natural follow-up that the paper does not run is to train a GPR on the same 25 Taguchi-selected points without active learning; that comparison would separate the benefit of sequential querying from the benefit of the nonlinear surrogate itself.
If the 15 test cases had been chosen by a documented random or stratified rule rather than an unspecified selection from the 125-point space, the reported margin might change; this is a testable condition for generalizing the result.
The single composite response $Y$ heavily weights penetration depth, so the claimed advantage is established for that weighted objective, not separately for height, width, or depth.
With an expected-improvement acquisition function, the active-learning loop could be turned into an inverse design tool that finds input parameters for a desired bead geometry, directly extending the paper's stated future work.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Before reading the paper, know one thing: the headline comparison (GPR beats Taguchi with fewer experiments) rests on a test set whose independence from the training pool is asserted in a sentence, not enforced in the methods. Section 2 says 15 combinations were selected from the 125-point space as test cases, and separately that the GPR iteratively acquired data from regions of high uncertainty. It never says the 15 were removed from the candidate pool. Since the design space is only 125 and GPR grabbed 20 points, random acquisition alone would expect about 2.4 overlap by chance. That's not a round-off issue. If any test point was ever a candidate for acquisition, the RMSE 1.00 vs 1.33 figure is not a fair comparison.

That said, the paper has real substance. It's an actual WAAM experimental campaign, not a simulation. The held-out evaluation is the right idea, the comparison is direct, and the active-learning procedure is standard (LHS start, uncertainty acquisition). The result isn't surprising — GPR with a nonlinear kernel should beat a linear main-effects Taguchi model on a nonlinear response. But the dataset and the clear comparison are useful for people doing process optimization in additive manufacturing.

The soft spots are mostly about missing detail rather than bad practice. No test-set selection protocol, no seed, no error bars from repeated runs, and no sensitivity analysis for the custom response variable Y. The Taguchi baseline is deliberately linear, which is fine, but the paper's language sometimes implies GPR is a superior method rather than a better model family. The most fixable issue is the test-set exclusion: the authors should either confirm the 15 cases were never in the acquisition pool, or provide the full acquisition history so readers can check.

One more thing: the code and data are promised in the supplementary but not accessible in the arXiv version. That matters here because the validation hinges on exactly what points were chosen.

I'd send this to review — the dataset is real and the question is practical — but I'd expect several rounds of revision before trusting the headline number. If the authors supply the acquisition log and show the test set was held out, the paper becomes a solid engineering contribution. If they don't, it's an unverifiable claim. Give it to a referee who will check the supplementary data.

Referee Report

4 major / 5 minor

Summary. The manuscript compares the Taguchi L25 orthogonal-array method with a Gaussian process regression (GPR) model that combines Latin hypercube sampling and uncertainty-based active learning for predicting a scalarized weld-bead geometry response in wire arc additive manufacturing. The authors report that, on 15 held-out test combinations, the final 20-point GPR model achieves RMSE 1.00 and R² 0.89, while the 25-point Taguchi main-effects model achieves RMSE 1.33 and R² 0.80. They further claim that GPR approaches Taguchi accuracy after seven experiments and surpasses it after seventeen. The paper presents this as evidence that active-learning GPR is both more accurate and more data-efficient than Taguchi for nonlinear process design.

Significance. If the reported comparison is valid, the paper provides a useful, concrete demonstration of active-learning GPR as a lower-cost alternative to Taguchi-based design of experiments in a physical manufacturing setting. The held-out evaluation design is a reasonable basic approach, and the explicit comparison in terms of number of experiments is directly relevant to experimental efficiency. The authors also make a clear effort toward reproducibility by promising Python code and data. However, the significance is currently limited by several unresolved methodological details: the test-set construction and its independence from the active-learning acquisition pool are not specified, no measurement uncertainty or replicate information is reported, and only a single active-learning trajectory is shown. These gaps make it impossible to assess whether the headline RMSE/R² advantage is robust or an artifact of experiment selection or noise.

major comments (4)

[Section 2, Figure 2(c)] The manuscript states that 15 combinations were selected from the 125-point design space as test cases, and then describes the GPR model as iteratively selecting new data points from regions with highest predictive uncertainty. It never states that these 15 test cases were excluded from the GPR candidate pool. Since GPR acquires 20 points from the same 125-point space, any overlap between the acquired points and the test set would invalidate the reported RMSE and R² values. The authors must provide the full list of acquired and test points, an explicit exclusion rule, or a reproducible protocol (e.g., a random seed with a documented split) so that the comparison is verifiable.
[Section 3, final GPR model] No measurement uncertainty is reported for the experimental bead-geometry measurements, and no replicate experiments are described. The final GPR model has a reported noise level of 1.70, and the RMSE difference between GPR (1.00) and Taguchi (1.33) is 0.33; without uncertainty quantification on the test-set predictions, this difference may not be statistically meaningful. The authors should report test-set prediction intervals, replicate measurements, or at least per-test errors so the reader can judge the significance of the claimed improvement.
[Section 2, GPR workflow] The active-learning procedure is underspecified. The acquisition function is described only as selecting points from regions of highest predictive uncertainty, and the GPR hyperparameter optimization is mentioned without giving bounds, kernels, or convergence criteria. Furthermore, the LHS initialization and active-learning iterations are evidently run once; no random seed or multiple-run variance is reported. Because LHS and acquisition are stochastic, a single run does not establish that the reported learning curve is representative. The authors should report multiple seeds or otherwise quantify run-to-run variability.
[Section 2, response parameter Y] The engineered response Y = log((D×10^10+W×10^6+H×10^2)/(D+W+H))×10 is introduced without justification for the specific weighting coefficients or a sensitivity analysis. Since all reported RMSE/R² comparisons are made on this scalarized response, the claims of GPR superiority are specific to this particular Y and may not carry over to the individual bead-geometry dimensions (height, width, depth). The authors should provide per-output predictions or argue that Y is a meaningful and robust scalarization for the intended process control.

minor comments (5)

[Abstract] The abstract contains a duplicated phrase: 'full factorial exploration. full factorial exploration.' This should be corrected.
[Introduction] The citation for Kim et al. appears as 'Kim et al. [? ]' in the text; a reference number is missing.
[Throughout] The abbreviation 'W AAM' contains an inconsistent space in several places (e.g., 'W AAM' vs. 'WAAM'); this should be made uniform.
[Section 3, Figure 3] The parity plot and error-distribution plots in Figure 3 do not show error bars or confidence intervals, which would help assess the practical significance of the RMSE/R² differences.
[Supplementary] The paper promises detailed predictions and Python code in Supporting Information, but the preprint as posted does not make these files accessible. Since the reproducibility of the active-learning and test-set procedures depends on them, the authors should ensure the code and data are available to reviewers and readers.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the GPR-versus-Taguchi comparison is empirically grounded, and the only self-citation is a non-load-bearing reference for a standard method.

full rationale

The central claim is empirically self-contained: the GPR model is trained on labeled experiments selected by LHS plus uncertainty-based active learning, and it is evaluated on 15 test cases that are not used to fit the model. Hyperparameters are optimized by 5-fold cross-validation on the training data, so the reported RMSE of 1.00 and R2 of 0.89 are test-set predictions, not fitted values renamed as predictions. The Taguchi baseline is likewise generated from its own L25 training points via the main-effects formula, so neither method's comparison reduces to its inputs by construction. The custom response Y is a constructed target, but it is applied identically to both methods and does not smuggle in the outcome. Reference [6] is a self-citation by co-author Batra for the active-learning/LHS strategy, but the same strategy is independently cited to Settles [10] and Deringer et al. [7], and the method is described explicitly in the text, so the self-citation is not load-bearing. A manuscript-level limitation is that Section 2 never states an explicit rule excluding the 15 test combinations from the GPR active-learning candidate pool; if any test point were ever acquired, the reported RMSE comparison would be invalid. That is a potential data-leakage or validity concern rather than a circular derivation, and the preprint's Supporting Information is not currently accessible to check it.

Assumptions & free parameters 5 free parameters · 6 assumptions · 0 invented entities

The central comparison rests on a hand-built response variable Y, fitted GPR hyperparameters, a linear Taguchi baseline, and an unspecified test-set selection. No new physical entities are introduced. The weakest entries are the ad hoc Y weights and the untested representativeness of the 15 test cases.

free parameters (5)

GPR kernel length scale = 2.68
Optimized by 5-fold cross-validation on training data; sets RBF smoothness and drives both predictions and uncertainty-based acquisition.
GPR noise level = 1.70
Optimized by 5-fold cross-validation; it directly scales the uncertainty estimates used in active learning.
Response Y weighting coefficients = 10^10, 10^6, 10^2 and factor 10
Hand-chosen constants in the Y = log(...) formulation in Section 2 prioritize penetration depth; no sensitivity analysis or physical basis is given.
Taguchi level effects (alpha_i, beta_j, gamma_k) = not reported
Fitted to 25 training runs for the main-effects model in Eq. (1); these are the baseline model's parameters.
Active learning budget = 5 initial LHS + 15 iterations (20 total)
The efficiency claim depends on this budget choice; no investigation of alternate budgets is reported.

assumptions (6)

domain assumption Taguchi main-effects model: predicted response equals grand mean plus additive level effects (Eq. 1)
The Taguchi baseline assumes no interactions and linear additive effects. The authors note Taguchi cannot capture nonlinearity, yet use this model for the comparison.
domain assumption Test set of 15 cases is representative and unbiased
Section 2 states 15 combinations were selected from the 125-point space but does not describe the selection procedure, random seed, or stratification.
domain assumption Single GPR active-learning run is representative
Only one LHS initial sample is described; no repeated runs or seed averaging are reported, so the trajectory in Figure 3(d) could be seed-dependent.
domain assumption RBF kernel with fitted hyperparameters is an adequate model class
GPR performance depends on kernel choice and stationarity; no comparison with other kernels or models is provided.
domain assumption Measurements are i.i.d. with negligible noise
No replicate experiments or measurement error bars are reported; ImageJ-based geometry measurements are treated as exact.
ad hoc to paper The engineered response Y is a valid scalarization of weld quality
The Y formula places arbitrary weights on penetration depth, width, and height; the comparison is only as meaningful as this target definition.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Enhancing Experimental Efficiency in Materials Design: A Comparative Study of Taguchi and Machine Learning Methods." pith.science (2026). https://pith.science/paper/IRWYG26U

@misc{pith2026250603910,
  author       = {Pith},
  title        = {Pith review of: Enhancing Experimental Efficiency in Materials Design: A Comparative Study of Taguchi and Machine Learning Methods},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IRWYG26U}},
  note         = {Machine review of arXiv:2506.03910}
}

read the original abstract

Materials design problems often require optimizing multiple variables, rendering full factorial exploration impractical. Design of experiment (DOE) methods, such as Taguchi technique, are commonly used to efficiently sample the design space but they inherently lack the ability to capture non-linear dependency of process variables. In this work, we demonstrate how machine learning (ML) methods can be used to overcome these limitations. We compare the performance of Taguchi method against an active learning based Gaussian process regression (GPR) model in a wire arc additive manufacturing (WAAM) process to accurately predict aspects of bead geometry, including penetration depth, bead width, and height. While Taguchi method utilized a three-factor, five-level L25 orthogonal array to suggest weld parameters, the GPR model used an uncertainty-based exploration acquisition function coupled with latin hypercube sampling for initial training data. Accuracy and efficiency of both models was evaluated on 15 test cases, with GPR outperforming Taguchi in both metrics. This work applies to broader materials processing domain requiring efficient exploration of complex parameters.

Figures

Figures reproduced from arXiv: 2506.03910 by the authors.

**Figure 1.** Overview of the WAAM design problem comparing the accuracy and efficiency of both models. The Taguchi method uses orthogonal sampling [3], while GPR-based method combines LHS [6] with uncertainty-based active learning [7]. A test set of 15 experiments is used for evaluation. Machine learning (ML) techniques, such as neural networks (NN) and Gaussian process regression (GPR), offer improved modeling of such nonlinear… view at source ↗

**Figure 2.** (a) WAAM setup for single weld bead deposition. (b) Input design space. (c) Overall workflow involving sample preparation, deposition, measurement, and experiment selection for Taguchi and GPR models. inputs. Five LHS samples were used for initial training to ensure uniform coverage and reduce selection bias [9]. To improve accuracy with minimal experiments, active learning was used. The GPR model iteratively select… view at source ↗

**Figure 3.** (a) Mean-of means plot (b) Parity plot comparing GPR model and Taguchi method against experimental results (c) Error distribution of Taguchi and GPR models.(d) Evolution of RMSE and R2 score of GPR model with each active learning iteration [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

12 extracted references · 8 canonical work pages

[1]

Y . Li, C. Su, J. Zhu, Comprehensive review of wire arc additive manufacturing: Hardware system, physical process, monitoring, property characterization, application and future prospects, Results Eng. 13 (2022) 100330. doi:10.1016/j.rineng.2021.100330

arXiv 2022
[2]

Y . Luo, J. Liu, H. Xu, C. Xiong, L. Liu, Regression modeling and process analysis of resistance spot welding on galvanized steel sheet, Mater. Des. 30 (7) (2009) 2547–2555. doi:10.1016/j.matdes.2008.09.031

work page doi:10.1016/j.matdes.2008.09.031 2009
[3]

D. C. Montgomery, Design and analysis of experiments, John Wiley & Sons, 2017

work page 2017
[4]

Dinovitzer, X

M. Dinovitzer, X. Chen, J. Laliberte, X. Huang, H. Frei, Effect of wire and arc additive manufacturing (waam) process parameters on bead geometry and microstructure, Addit. Manuf. 26 (2019) 138–146. doi:10.1016/j.addma.2018.12.013. 6

work page doi:10.1016/j.addma.2018.12.013 2019
[5]

Lin, J.-C

H.-L. Lin, J.-C. Yan, Optimization of weld bead geometry in the activated gma welding process via a grey-based taguchi method, J. Mech. Sci. Technol. 28 (2014) 3249–3254. doi:10.1007/s12206-014-0735-9

work page doi:10.1007/s12206-014-0735-9 2014
[6]

Batra, L

R. Batra, L. Song, R. Ramprasad, Emerging materials intelligence ecosystems propelled by machine learning, Nat. Rev. Mater. 6 (8) (2021) 655–678. doi:10.1038/s41578-020-00255-y

doi:10.1038/s41578-020-00255-y 2021
[7]

V . L. Deringer, A. P. Bart´ok, N. Bernstein, D. M. Wilkins, M. Ceriotti, G. Cs´anyi, Gaussian process regression for materials and molecules, Chem. Rev. 121 (16) (2021) 10073–10141. doi:10.1021/acs.chemrev.1c00022

doi:10.1021/acs.chemrev.1c00022 2021
[8]

Z. Wang, S. Zimmer-Chevret, F. L ´eonard, G. Abba, Prediction of bead geometry with consideration of interlayer temperature e ffect for cmt-based wire-arc additive manufacturing, World Weld. 65 (12) (2021) 2255–2266. doi:10.1007/s40194-021-01192-2

work page doi:10.1007/s40194-021-01192-2 2021

Show all 12 references

[9]

R. L. Iman, Latin hypercube sampling, in: N. Balakrishnan (Ed.), Methods and Applications of Statistics in the Atmospheric and Earth Sciences, Wiley, Hoboken, NJ, 2012, Ch. 15, pp. 140–145

2012
[10]

Settles, Active learning literature survey, Technical Report 1648, University of Wisconsin-Madison, Dept

B. Settles, Active learning literature survey, Technical Report 1648, University of Wisconsin-Madison, Dept. of Computer Sciences (2009)

2009
[11]

Geron, Hands-On Machine Learning with Scikit-Learn, Keras, and TensorFlow: Concepts, Tools, and Techniques to Build Intelligent Systems, 2nd Edition, O’Reilly Media, Inc., 2019

A. Geron, Hands-On Machine Learning with Scikit-Learn, Keras, and TensorFlow: Concepts, Tools, and Techniques to Build Intelligent Systems, 2nd Edition, O’Reilly Media, Inc., 2019

2019
[12]

Mattera, G

G. Mattera, G. Piscopo, M. Longobardi, M. Giacalone, L. Nele, Improving the interpretability of data-driven models for additive manufacturing processes using clusterwise regression, J. Math. 12 (16) (2024). doi:10.3390/math12162559. 7

2024 doi

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

Y . Li, C. Su, J. Zhu, Comprehensive review of wire arc additive manufacturing: Hardware system, physical process, monitoring, property characterization, application and future prospects, Results Eng. 13 (2022) 100330. doi:10.1016/j.rineng.2021.100330

arXiv 2022

[2] [2]

Y . Luo, J. Liu, H. Xu, C. Xiong, L. Liu, Regression modeling and process analysis of resistance spot welding on galvanized steel sheet, Mater. Des. 30 (7) (2009) 2547–2555. doi:10.1016/j.matdes.2008.09.031

work page doi:10.1016/j.matdes.2008.09.031 2009

[3] [3]

D. C. Montgomery, Design and analysis of experiments, John Wiley & Sons, 2017

work page 2017

[4] [4]

Dinovitzer, X

M. Dinovitzer, X. Chen, J. Laliberte, X. Huang, H. Frei, Effect of wire and arc additive manufacturing (waam) process parameters on bead geometry and microstructure, Addit. Manuf. 26 (2019) 138–146. doi:10.1016/j.addma.2018.12.013. 6

work page doi:10.1016/j.addma.2018.12.013 2019

[5] [5]

Lin, J.-C

H.-L. Lin, J.-C. Yan, Optimization of weld bead geometry in the activated gma welding process via a grey-based taguchi method, J. Mech. Sci. Technol. 28 (2014) 3249–3254. doi:10.1007/s12206-014-0735-9

work page doi:10.1007/s12206-014-0735-9 2014

[6] [6]

Batra, L

R. Batra, L. Song, R. Ramprasad, Emerging materials intelligence ecosystems propelled by machine learning, Nat. Rev. Mater. 6 (8) (2021) 655–678. doi:10.1038/s41578-020-00255-y

doi:10.1038/s41578-020-00255-y 2021

[7] [7]

V . L. Deringer, A. P. Bart´ok, N. Bernstein, D. M. Wilkins, M. Ceriotti, G. Cs´anyi, Gaussian process regression for materials and molecules, Chem. Rev. 121 (16) (2021) 10073–10141. doi:10.1021/acs.chemrev.1c00022

doi:10.1021/acs.chemrev.1c00022 2021

[8] [8]

Z. Wang, S. Zimmer-Chevret, F. L ´eonard, G. Abba, Prediction of bead geometry with consideration of interlayer temperature e ffect for cmt-based wire-arc additive manufacturing, World Weld. 65 (12) (2021) 2255–2266. doi:10.1007/s40194-021-01192-2

work page doi:10.1007/s40194-021-01192-2 2021