First Contact: Data-driven Friction-Stir Process Control

David Garcia; Ethan King; James Koch; Keerti Kappagantula; Ken Ross; Megan Ebers; Woongjo Choi

REVIEW 4 major objections 6 minor 11 references

First Contact: Data-driven Friction-Stir Process Control

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read A trained neural lumped-parameter ODE model, fit to seven plunge experiments, generates cubic power profiles that drive a real friction-stir tool to within 1% of target handoff temperatures in six validation runs.

desk verdict A genuine hardware proof-of-concept for data-driven FSP temperature control, but the unverified equivalence between commanded and delivered power is the load-bearing gap. read the letter →

arxiv 2507.03177 v1 pith:3A6IPXO6 submitted 2025-07-03 eess.SY cs.LGcs.SY

James Koch , Ethan King , WoongJo Choi , Megan Ebers , David Garcia , Ken Ross , Keerti Kappagantula This is my paper · ORCID

classification eess.SYcs.LGcs.SY

keywords frictionstirprocessingopen-loopcontrolneuralordinarydifferentialequationslumped-parametermodelsystemidentificationsetpointtooltemperaturepowerprofileoptimization

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

This paper reports that a small data-driven model of tool temperature can be used as a controller for friction-stir processing. The authors train a neural lumped-parameter ODE—a first-law energy balance with a neural network standing in for the unknown power-to-heat conversion factor—on seven plunge experiments in 316L stainless steel. They then freeze the model and optimize cubic power profiles to reach target tool temperatures, running those profiles on real hardware. In six validation experiments at 750 and 775 degrees Celsius, the final handoff temperature was within 1% of setpoint, even though neither setpoint appeared in the training data. If this holds in broader use, it would replace hand-tuned plunge schedules with a short, interpretable training-and-optimization pipeline.

What carries the argument

The central object is the neural lumped-parameter differential equation dT/dt = (1/C)(eta(T,P;$\theta$)P(t) - h(T - T_sink)), where a feedforward network with one hidden layer of 15 sigmoid units replaces the unknown conversion efficiency eta. The network output is constrained to (0,1), the heat-loss coefficient h and capacitance C are learned alongside it, and automatic differentiation through the ODE solver provides gradients for both system identification and, with frozen weights, trajectory optimization. The control signal it produces is a cubic polynomial P(t;phi) = phi_1 + phi_2 t + phi_3 $t^{2}$ + phi_4 $t^{3}$, with phi_1 fixed at the idle power draw; this is exactly the format the machine's power-control software accepts. Training tunes the unknown physics, and freezing turns the model into a differentiable simulator whose inputs can be optimized for setpoint accuracy, energy use, smoothness, and overshoot.

What would settle it

Instrument the friction-stir machine to log realized spindle power during the plunge and compare it point-by-point with the commanded cubic polynomial. If the mismatch is large enough that the trained model would predict a different final temperature than observed, then the 1% handoff results cannot be attributed to the model alone, and the open-loop control claim would need to be revised.

Watch

Extended reading notes

Core claim

The central claim is that a first-law lumped-capacitance temperature model, in which the unknown conversion factor from machine power to internal heating is replaced by a small sigmoid-output neural network, can be fitted by gradient descent through an ODE solver and then used, with weights frozen, to design open-loop control inputs that achieve set-point temperatures on an actual friction-stir processing machine. The control input is constrained to the third-order polynomial power profile the machine's built-in controller expects, with the constant term fixed at the measured idle power draw. Across two validation batches, all six runs reached final handoff temperatures at or below 1% of the 750 or 775 degree Celsius setpoints. Mean absolute percent errors of roughly 7 to 20 percent during the plunge are dominated by a time lag in the first about 20 seconds after tool contact, and the authors note that model and experiment align after that. The result is offered as validation that interpretable, data-enabled system dynamics models can carry out real-world process control tasks.

Load-bearing premise

The load-bearing assumption is that the machine's internal power controller delivers the commanded cubic power profile to the workpiece accurately enough that the model input matches the physics; the paper reports no direct measurement of delivered power or tracking error.

Editorial extensions

If this is right

All six validation experiments reached handoff within 1% of setpoint, including two setpoints absent from the training data, so the method transfers to unseen setpoints without manual retuning.
Changing a single smoothing hyperparameter produced distinct fast and slow profiles to the same setpoint, indicating the framework can shape time-to-temperature.
The differentiable model plus trajectory optimization gives a direct path to closed-loop feedback control once hardware supports real-time sampling and control.
The same system identification pipeline should apply to other alloys and processing conditions by retraining or fine-tuning the neural conversion factor.

Reading between the lines

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

The 1% end-state criterion is what is validated; the paper's own error metrics show the model is not yet accurate about the early transient, so any claim about defect formation during the first seconds of plunge is not established.
Because delivered power is never measured, the neural network may have implicitly absorbed the machine's internal controller behavior; the optimized profiles may need re-validation on a machine with different power-tracking characteristics.
A direct next experiment would be to rerun one of the validated profiles after intentionally changing the machine's power-controller gains; if handoff accuracy persists, the model is robust to actuator details, and if not, the controller tracking is the true load-bearing element.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

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

Quick take: this is an honest, incremental hardware validation of a method the authors proposed last year. The new thing is not the UDE/Neural ODE machinery—that's standard and they own it—but the fact that a trained lumped model can generate cubic power profiles that, when loaded into an FSP machine, bring tool temperature to two unseen setpoints within 1% at handoff. That is a real result and it is directly measured by a thermocouple. The paper also acknowledges the obvious blemish: model-experiment mismatch in the first 20 seconds of plunge, which inflates MAPE but not the handoff error. Credit where due: they tested extrapolation to setpoints outside the training range, they report both fast and slow profiles in batch 2, and the loss hyperparameters are explicit.

The soft spot is the one the stress-test flags, and it is load-bearing. The model is trained on realized power, but deployment commands a cubic profile to the machine's internal PID, which only 'attempts' to reach that power. No measured power, spindle torque, or tracking error is reported for any validation run. If the PID lags or saturates, the observed temperature trajectory is not the response to the profile the model generated, and the within-1% handoff temperatures cannot be cleanly attributed to the model. The authors need to show the realized power trace, or at least quantify the tracking error, before the central claim is robust. The validation is also thin on statistics: only two unique setpoints, no repeated trials, no error bars, no baseline comparison to hand-tuned power ramps, and no code or data. Minor, but worth saying.

On balance the central claim—open-loop setpoint control from a data-driven lumped model—is plausible and the experiment is a genuine first. The gaps are fixable and the paper is honest about the model's early-time deficiency. I'd send it to peer review, but I'd ask the reviewers to demand the realized-power data. This is a paper for the FSP and process-control community, and for applied SciML people who care about deployment. It's not a breakthrough, but it's a concrete proof-of-concept that deserves a serious referee.

Referee Report

4 major / 6 minor

Summary. The manuscript proposes and experimentally tests an open-loop, model-based controller for the plunge phase of friction-stir processing (FSP). The authors train a neural lumped-parameter ODE (Eq. 1) on temperature time series from seven FSP runs, then freeze that model and optimize the coefficients of a cubic power profile (Eq. 4) to drive the modeled tool temperature to a desired setpoint while penalizing energy, roughness, and overshoot. The optimized profiles are loaded into the machine's built-in power controller and evaluated in six experiments at two setpoints (750 and 775 °C). The paper reports that the measured tool temperature at the power-to-temperature handoff was within 1% of setpoint in all six runs.

Significance. If the claimed results are robust, this is a worthwhile proof-of-concept: it demonstrates that a physically structured data-driven model can generate open-loop power profiles for FSP without manual tuning, and the method is transparent and inexpensive to train. The clarity of the model description and the use of standard, reproducible numerical tools (Julia's DifferentialEquations.jl and DiffEqFlux.jl) are strengths. However, the validation is currently too limited to establish the general capability claimed in the title and abstract.

major comments (4)

[§2.4, with Figs. 2b and 4-9] The controller commands a polynomial power profile to the machine's internal PID, but the model in Eq. (1) was trained on realized power ("given actual power input," Fig. 2b). The paper does not report the actual power delivered during the six validation runs, nor the PID tracking error, spindle speed, or torque. If the realized power deviates from the commanded cubic profile (e.g., due to PID lag or the 1 kW idle offset), the measured temperature trajectories cannot be attributed to the model-generated profile. Please provide measured power (or acceptable proxies) and quantify tracking error; if these data are unavailable, the central claim must be explicitly limited to commanded, not delivered, power.
[§3.1-3.2] The validation is thin: only two previously unseen setpoints (775 and 750 °C) appear in Batch 1, and Batch 2 repeats those same two setpoints with different speed profiles. Each condition was run once, so there are no error bars or repeated trials, and no baseline comparison (e.g., hand-tuned linear ramps or the machine's default temperature control) is reported. The "within 1%" claim therefore rests on single measurements per condition. I ask for at least three repetitions per condition and a comparison to a standard or manual baseline to support the claim that the data-driven profiles provide value.
[Eq. (3), L_smoothing] The smoothing penalty is defined as a sum of first differences P(t^{k+1}) - P(t^k) without an absolute value or a square. As written, this term can be negative and does not penalize roughness as the text claims; it would instead reward a monotonically decreasing profile. Please correct the equation (likely |ΔP| or (ΔP)^2) and confirm what was actually implemented in the optimization, since the loss is load-bearing for the reported control profiles.
[§3, handoff metric] The primary performance metric is the temperature at the moment the machine automatically switches from power control to temperature control. Because the switch is triggered by reaching the setpoint, the "handoff temperature" is partly determined by the machine's own threshold logic, which makes the "within 1%" claim less informative than a metric like time-to-setpoint, overshoot, or the full temperature trajectory error. The paper reports Max T in Table 1, but it should explicitly report the time-to-handoff and overshoot for each run, and the model-experiment error over the entire plunge, not only at the handoff.

minor comments (6)

[§4] The term "model predictive control" is used to describe the offline trajectory optimization of Section 2.3, but true MPC employs closed-loop receding-horizon feedback; please use "open-loop optimal control" or "trajectory optimization" to avoid confusion.
[Eq. (3)] The overshoot term uses relu but does not define it; please define relu(x)=max(0,x).
[Table 1] The meaning of "Max T" is unclear (maximum temperature during the plunge, after handoff, or including the traverse?); please define it in the caption.
[§2.2] "an scheme" should be "a scheme".
[§2.2-2.3] The paper does not report the number of data points per experimental run or the time step used in the ODE solver; please add these details for reproducibility.
[Table 1] The MAPE values up to 20.5% are attributed to time delays in the initial contact phase; a brief discussion of whether these delays arise from PID tracking error or from model error would help the reader assess the controller's robustness.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the control profiles are optimized against a fitted model, and the validation is an independent hardware test on setpoints outside the training distribution in Batch #1.

full rationale

The central derivation chain is system identification (Eq. 2) followed by open-loop trajectory optimization (Eq. 3) and hardware validation. The model is fitted to temperature time series from hand-tuned linear power profiles; the control optimization then freezes the fitted parameters and minimizes a setpoint loss with respect to the polynomial power coefficients. The resulting profile is deployed on the FSP machine. The final handoff temperatures are experimental measurements, not quantities recovered from the optimization objective, so the within-1% claim is not equivalent to the training objective by construction. Batch #1 explicitly targets setpoints (775C and 750C) that were absent from the training data, making the hardware test an extrapolation. Batch #2 retrains on the added Batch #1 data, but the new fast/slow profiles are distinct control inputs, so the validation still tests the model on new inputs. The only citation to prior work by the same authors, Ref. [4], supplies the model form and dataset details, but the system ID and control optimizations are re-run in the present paper, so the citation is not load-bearing circularity. A genuine limitation is that the machine's internal PID tracking of the commanded cubic profile is not measured, so realized power is assumed to equal commanded power; this is an experimental-validation gap (correctness risk), not a circularity. No step reduces to its own inputs by definition.

Assumptions & free parameters 4 free parameters · 7 assumptions · 0 invented entities

The model is a semi-empirical fit: the ODE skeleton is physics-based, but the neural network η is a flexible function approximator, and the hyperparameters and polynomial control form are user choices. No new physical entities are introduced.

free parameters (4)

Thermal capacitance C = not reported
Appears in Eq. 1; tuned by system ID to match seven training runs.
Heat transfer coefficient h = not reported
Tuned in Eq. 1; controls heat loss to sink.
Neural network weights θ = not reported (61 weights for a 2-15-1 MLP)
Fitted via Adam then BFGS on the temperature reconstruction loss (Eq. 2).
Loss hyperparameters λ1-4 = (1, 1e3, 1e6, 1e3) fast; (1, 1e3, 1e7, 1e3) slow
Hand-chosen in Eq. 3; they shape the optimized power profile and define fast versus slow.

assumptions (7)

standard math First Law of Thermodynamics energy balance
The ODE in Eq. 1 is derived from dE/dt = Qdot - Wdot, invoked in Section 2.1.
domain assumption Lumped capacitance with uniform temperature
The model assumes a single thermal capacitance C and uniform temperature; stated in Section 2.1.
domain assumption Linear heat transfer to a constant sink
Eq. 1 uses h(T - Tsink) for heat loss; no radiative or nonlinear losses are included.
ad hoc to paper Neural network η(T,P) with sigmoid output in (0,1) can represent the true power conversion efficiency
The unknown conversion factor is replaced by a 2-15-1 MLP with sigmoid activations, Section 2.1. This is a flexible function approximation with no physical justification beyond the ODE structure.
ad hoc to paper Power profile restricted to a third-order polynomial
The machine's built-in control software expects a third-order polynomial power profile (Section 2.4), so the optimal control is constrained to this form.
domain assumption The machine's PID controller tracks commanded power
Section 2.4 states the machine 'attempts to reach a desired power level' via a tuned PID, but no measured power or tracking error is reported.
domain assumption Thermocouple temperature in tool shoulder is the controlled temperature
All experimental temperatures are from a thermocouple in the tool shoulder, Section 2.4; the model treats this as the system state.

how reviews work

0 comments

Cite this review

Pith. "Pith review of First Contact: Data-driven Friction-Stir Process Control." pith.science (2026). https://pith.science/paper/3A6IPXO6

@misc{pith2026250703177,
  author       = {Pith},
  title        = {Pith review of: First Contact: Data-driven Friction-Stir Process Control},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3A6IPXO6}},
  note         = {Machine review of arXiv:2507.03177}
}

read the original abstract

This study validates the use of Neural Lumped Parameter Differential Equations for open-loop setpoint control of the plunge sequence in Friction Stir Processing (FSP). The approach integrates a data-driven framework with classical heat transfer techniques to predict tool temperatures, informing control strategies. By utilizing a trained Neural Lumped Parameter Differential Equation model, we translate theoretical predictions into practical set-point control, facilitating rapid attainment of desired tool temperatures and ensuring consistent thermomechanical states during FSP. This study covers the design, implementation, and experimental validation of our control approach, establishing a foundation for efficient, adaptive FSP operations.

Figures

Figures reproduced from arXiv: 2507.03177 by the authors.

**Figure 1.** Friction-stir processing (FSP) uses a rotating tool to heat and plastically deform material of a workpiece. In (a), shown are notional time histories for power and tool temperature during the plunge sequence of FSP. After reaching a desired temperature set point, the plunge sequence concludes and the tool traverse in the sample commences. Here, we seek to quickly obtain a set point temperature through data-driven op… view at source ↗

**Figure 2.** Training loss history for the system ID task is shown in (a). An example trajectory reconstruction [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗

**Figure 3.** The system dynamics model of Eq. 1 contains a neural network stand-in for the unknown conversion [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

Figures from the paper (3 more)

**Figure 4.** Figure 4: Batch #1. 775C set point [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]

**Figure 6.** Figure 6: Batch #2. 775C set point (fast) [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]

**Figure 8.** Figure 8: Batch #2. 775C set point (slow) [PITH_FULL_IMAGE:figures/full_fig_p008_8.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

11 extracted references · 6 canonical work pages

[1]

write newline

" write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in ":" * " " * FUNCTION f...
[2]

write newline

" write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in ":" * " " * FUNCTION f...
[3]

write newline

" write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in ":" * " " * FUNCTION f...
[4]

R. S. Mishra, Z. Ma, Friction stir welding and processing, Materials science and engineering: R: reports 50 (1-2) (2005) 1--78

work page 2005
[5]

Heidarzadeh, S

A. Heidarzadeh, S. Mironov, R. Kaibyshev, G. C am, A. Simar, A. Gerlich, F. Khodabakhshi, A. Mostafaei, D. P. Field, J. D. Robson, et al., Friction stir welding/processing of metals and alloys: a comprehensive review on microstructural evolution, Progress in Materials Science 117 (2021) 100752

work page 2021
[6]

P. Kah, R. Rajan, J. Martikainen, R. Suoranta, Investigation of weld defects in friction-stir welding and fusion welding of aluminium alloys, International Journal of Mechanical and Materials Engineering 10 (2015) 1--10

work page 2015
[7]

J. Koch, W. Choi, E. King, D. Garcia, H. Das, T. Wang, K. Ross, K. Kappagantula, Neural lumped parameter differential equations with application in friction-stir processing, Journal of Intelligent Manufacturing (2024) 1--11

work page 2024
[8]

Rackauckas, Y

C. Rackauckas, Y. Ma, J. Martensen, C. Warner, K. Zubov, R. Supekar, D. Skinner, A. Ramadhan, A. Edelman, Universal differential equations for scientific machine learning, arXiv preprint arXiv:2001.04385 (2020)

arXiv 2020

Show all 11 references

[9]

R. T. Chen, Y. Rubanova, J. Bettencourt, D. K. Duvenaud, Neural ordinary differential equations, Advances in neural information processing systems 31 (2018)

2018
[10]

Drgo n a, K

J. Drgo n a, K. Ki s , A. Tuor, D. Vrabie, M. Klau c o, Differentiable predictive control: Deep learning alternative to explicit model predictive control for unknown nonlinear systems, Journal of Process Control 116 (2022) 80--92

2022
[11]

Rackauckas, Q

C. Rackauckas, Q. Nie, Differentialequations.jl--a performant and feature-rich ecosystem for solving differential equations in julia, Journal of Open Research Software 5 (1) (2017)

2017

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [1]

write newline

" write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in ":" * " " * FUNCTION f...

[2] [2]

write newline

" write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in ":" * " " * FUNCTION f...

[3] [3]

write newline

" write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in ":" * " " * FUNCTION f...

[4] [4]

R. S. Mishra, Z. Ma, Friction stir welding and processing, Materials science and engineering: R: reports 50 (1-2) (2005) 1--78

work page 2005

[5] [5]

Heidarzadeh, S

A. Heidarzadeh, S. Mironov, R. Kaibyshev, G. C am, A. Simar, A. Gerlich, F. Khodabakhshi, A. Mostafaei, D. P. Field, J. D. Robson, et al., Friction stir welding/processing of metals and alloys: a comprehensive review on microstructural evolution, Progress in Materials Science 117 (2021) 100752

work page 2021

[6] [6]

P. Kah, R. Rajan, J. Martikainen, R. Suoranta, Investigation of weld defects in friction-stir welding and fusion welding of aluminium alloys, International Journal of Mechanical and Materials Engineering 10 (2015) 1--10

work page 2015

[7] [7]

J. Koch, W. Choi, E. King, D. Garcia, H. Das, T. Wang, K. Ross, K. Kappagantula, Neural lumped parameter differential equations with application in friction-stir processing, Journal of Intelligent Manufacturing (2024) 1--11

work page 2024

[8] [8]

Rackauckas, Y

C. Rackauckas, Y. Ma, J. Martensen, C. Warner, K. Zubov, R. Supekar, D. Skinner, A. Ramadhan, A. Edelman, Universal differential equations for scientific machine learning, arXiv preprint arXiv:2001.04385 (2020)

arXiv 2020