{"id":"00bd92ee-38f7-40a5-92a2-da0ff9574b1e","arxiv_id":"2507.03177","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A data-driven lumped-parameter ODE with a neural-network efficiency term can synthesize open-loop power ramps that bring friction-stir tool temperature to within 1% of unseen setpoints.","lead":"Researchers trained a neural-network-enhanced heat model on temperature data from seven friction-stir processing runs, then used it to design power profiles that hit new target tool temperatures on real hardware. The tool temperature at the moment of process handoff landed within 1% of the target in all six test trials.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The experimental validation does not establish that the optimized power profiles were actually delivered: the machine's PID tracking of the commanded cubic profiles is never quantified, so realized power may differ from the model input.","rationale":"The reader's verdict is CONDITIONAL, and the weakest assumption identified is precisely the gap I find most load-bearing: the training/control input mismatch. Section 2.2 fits Eq. 1 to realized power, and Fig. 2b explicitly says 'given actual power input.' Section 2.4 then deploys a commanded polynomial through a PID that 'attempts to reach a desired power level.' No realized-power data are reported for any of the six validation experiments. Without that data, the central claim—that the model-generated cubic profiles drive the temperature to setpoint—is not actually tested; a different realized power profile could have produced the same temperature outcome. A secondary concern is that the handoff-based endpoint weakens the reported success metric, since the machine switches to temperature control when the setpoint is detected, so the within-1% final temperature is not fully independent of the experimental protocol. Still, the temperature trajectories and low maximum temperatures give some support. The power-tracking gap is sufficient to keep the paper at CONDITIONAL: the method is plausible and the experimental results are encouraging, but the missing actuator verification should be supplied before stronger claims are made.","tokens_in":6086,"tokens_out":8544,"duration_ms":110080,"concrete_test":"For all six validation runs, extract the machine-logged realized power (or spindle power/torque) time series and compare them with the commanded third-order polynomials listed in Table 1 over the full plunge until handoff. Quantify the tracking error as a time-resolved difference, e.g., RMSE and max absolute error relative to the commanded range. If the error is small (e.g., less than 5% of the commanded range), the concern is resolved; if it is large, rerun or reinterpret the validation, since realized power, not the command, is what the model must explain.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that the optimized cubic power profiles P(t; φ) of Eq. 4 drive the FSP tool to setpoint temperatures. This requires that the power actually delivered to the workpiece matches the power used as the model input. However, the system-identification model in Eq. 1 is fit to realized power (Section 2.2, Fig. 2b: 'given actual power input'), while deployment commands a profile to the machine's internal PID, which Section 2.4 says only 'attempts to reach a desired power level.' No measured power, tracking error, or spindle/torque data are reported for any of the six validation runs. If realized power deviates from the commanded polynomial because of PID lag, saturation, or the 1 kW idle offset, the observed temperature trajectories are not the response to the model-generated profile, and the within-1% handoff temperatures cannot be attributed to the model. The paper is otherwise a plausible proof-of-concept, but this unstated equivalence between commanded and realized power is load-bearing.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","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.","tokens_in":6301,"tokens_out":8474,"duration_ms":97886,"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":[{"comment":"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.","section":"§2.4, with Figs. 2b and 4-9"},{"comment":"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.","section":"§3.1-3.2"},{"comment":"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.","section":"Eq. (3), L_smoothing"},{"comment":"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.","section":"§3, handoff metric"}],"minor_comments":[{"comment":"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.","section":"§4"},{"comment":"The overshoot term uses relu but does not define it; please define relu(x)=max(0,x).","section":"Eq. (3)"},{"comment":"The meaning of \"Max T\" is unclear (maximum temperature during the plunge, after handoff, or including the traverse?); please define it in the caption.","section":"Table 1"},{"comment":"\"an scheme\" should be \"a scheme\".","section":"§2.2"},{"comment":"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.","section":"§2.2-2.3"},{"comment":"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.","section":"Table 1"}],"recommendation":"major_revision","confidential_remarks":"This is a short proof-of-concept that could become acceptable after a significant revision. The main blocker is the absence of measured power data in the validation runs; without it, the paper cannot substantiate that the optimized profile was actually delivered. The smoothing-loss equation also appears to be incorrect as written. The journal may also want to consider whether the number of experiments (six, no repeats) justifies the title's \"First Contact\" framing, but the direction is promising."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","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.\n\nThe 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.\n\nOn 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.","headline":"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.","tokens_in":6825,"tokens_out":2103,"would_cite":false,"duration_ms":24981,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"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.","keywords":["friction stir processing","open-loop control","neural ordinary differential equations","lumped-parameter model","system identification","setpoint control","tool temperature","power profile optimization"],"falsifier":"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.","tokens_in":5869,"feed_emoji":"🔥","tokens_out":6058,"duration_ms":67513,"temperature":0.7,"pith_summary":"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.","feed_headline":"Data-driven power profiles land friction-stir setpoints within 1%","feed_subtitle":"A seven-run training set becomes cubic power profiles that steer real tool temperatures to 750 or 775°C.","key_machinery":"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.","core_discovery":"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.","pith_inferences":["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."],"forward_implications":["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."],"supporting_citations":[{"why":"Introduces the neural lumped-parameter differential equation and supplies the seven-experiment temperature dataset this work trains on.","marker":"[4]"},{"why":"Establishes the universal differential equation paradigm of augmenting physics-based ODEs with trainable neural networks, the basis for the system identification step.","marker":"[5]"},{"why":"Provides neural ODE training by automatic differentiation through numerical integration, which makes both system identification and trajectory optimization possible.","marker":"[6]"},{"why":"Cited as the differentiable predictive control formulation that the closed-loop extension of this open-loop framework would build on.","marker":"[7]"}],"fun_headline_variants":["Neural differential equations steer friction-stir to 1% setpoint","Data-driven control gets friction-stir temps within 1%","Neural-ODE power profiles hit friction-stir setpoints","Machine-learned power shapes hit friction-stir temperature targets","Seven-run training set yields friction-stir control within 1%"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"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.","fun_headline_variants_meta":{"raw":{"variants":["Neural differential equations steer friction-stir to 1% setpoint","Data-driven control gets friction-stir temps within 1%","Neural-ODE power profiles hit friction-stir setpoints","Machine-learned power shapes hit friction-stir temperature targets","Seven-run training set yields friction-stir control within 1%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001081,"raw_usage":{"total_tokens":4469,"prompt_tokens":839,"completion_tokens":3630,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":455,"completion_tokens_details":{"reasoning_tokens":3542}},"tokens_in":455,"tokens_out":3630,"duration_ms":29027,"temperature":1.0,"reasoning_tokens":3542,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T20:16:18.712636+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"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.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the neural lumped-parameter differential equation and supplies the seven-experiment temperature dataset this work trains on."},{"cited_title":"Heidarzadeh, S","cited_arxiv_id":null,"evidence_quote":"Establishes the universal differential equation paradigm of augmenting physics-based ODEs with trainable neural networks, the basis for the system identification step."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides neural ODE training by automatic differentiation through numerical integration, which makes both system identification and trajectory optimization possible."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Cited as the differentiable predictive control formulation that the closed-loop extension of this open-loop framework would build on."}],"review_version":1}