{"id":"41c69cf5-9428-4bc6-9856-d5630ab3b17c","arxiv_id":"2506.08157","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A neural network trained on synthetic data estimates thrust from in-situ sensors, and an RCAC-based adaptive PID regulates the thrust of a simulated solid fuel ramjet with a variable-geometry inlet.","lead":"This paper combines a neural network thrust estimator with an adaptive controller to regulate solid fuel ramjet thrust using limited onboard sensors, tested in simulations of a quasi-one-dimensional SFRJ model. The approach is relevant to propulsion control, where direct in-flight thrust measurement is difficult.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper never states whether the closed-loop simulations use the ANN-predicted thrust or the true model thrust as feedback in Eq. (11); without this, the central claim that the NN+RCAC framework regulates thrust with in-situ data is not verifiable.","rationale":"The reader's weakest assumption was the fidelity of the quasi-1D SFRJ model. That is a legitimate external validity concern, but it is not load-bearing for the claim as worded, which is about regulating 'the thrust produced by the SFRJ model.' The load-bearing question is whether the simulations actually exercised the architecture in Fig. 4, where the ANN is in the feedback path. The ambiguity arises from Section IV's silence on the feedback signal; the conclusion's statement could be read as an assumption rather than a description of implementation. This is the kind of missing detail that a single numerical experiment can settle. The paper earns credit for a clear model description, a sensitivity study on ANN hyperparameters, and a demonstration of RCAC on a nontrivial plant model, but the presented evidence does not yet establish the combined framework as claimed. Conditional acceptance is appropriate pending the proposed verification.","tokens_in":10774,"tokens_out":6176,"duration_ms":72292,"concrete_test":"Re-run the step command scenario of Section IV-A.1 with the same RCAC hyperparameters twice: once with yk equal to the true model thrust T (Eq. 8), and once with yk equal to the neural-network estimate T_hat given by the trained ANN, while recording the true plant thrust in both cases. If the true thrust in the ANN-feedback run tracks the 100 N command after transient (e.g., absolute tracking error below 1 N) and stays stable over the full burn, the claim is supported; otherwise, the claimed NN+RCAC framework is not demonstrated.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The load-bearing issue is internal to the paper's own demonstration. In Section III, the authors describe two scenarios: an ideal one where the feedback yk is the true thrust, and a practical one where yk is the ANN estimate of thrust from in-situ measurements. Section IV presents 'command following using in situ measurements,' but the text and figures never state which of the two signals was used as yk in the RCAC update (Eq. 11). The conclusion asserts that simulations used the neural network predictions as the proxy for thrust. If the simulations actually used the true model thrust T from Eq. (8), then the central claim that the 'combined neural network and learning control framework' regulates thrust is not demonstrated—the results would only show RCAC with direct thrust feedback. If the simulations did use the ANN output, then two omissions weaken the claim: Section II-B explicitly omits the testing loss ('not shown for brevity'), and there is no closed-loop comparison showing that the ANN-based feedback achieves tracking comparable to direct thrust feedback. Because the plant and training data come from the same quasi-1D model, the ANN error is likely small in-distribution, but any unexamined bias could interact with the integral term in Eq. (11) and cause offset or instability. The paper as written does not permit the reader to verify which feedback path was simulated.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a framework for thrust monitoring and regulation of a quasi-one-dimensional solid fuel ramjet (SFRJ) model using a neural network (ANN) thrust estimator coupled with retrospective cost adaptive control (RCAC). The ANN maps in-situ measurements (cowl radius, combustor total pressure, CO mole fraction, altitude) to thrust, and is trained on 125,000 synthetic samples generated from the same quasi-1D model used as the plant. An adaptive PID controller, whose gains are updated online by RCAC, modulates the cowl radius to follow commanded thrust profiles. The paper presents simulations for step, doublet, ramp, and harmonic commands, and includes sensitivity studies for ANN hyperparameters and RCAC hyperparameters. The central claim is that the combined ANN+RCAC framework can regulate thrust using limited in-situ data, based on the numerical simulations.","tokens_in":11028,"tokens_out":4746,"duration_ms":59130,"significance":"If the central claim is fully supported, the paper offers a useful direction for SFRJ control without direct thrust measurement, which is difficult in flight. The strengths of the manuscript are its detailed description of the quasi-1D model and data-generation procedure, the explicit ANN architecture and training setup, the complete RCAC update equations, and the hyperparameter sensitivity studies. These elements make the synthetic-data portion of the work largely reproducible. The main weakness is that the simulation section does not unambiguously identify which feedback signal is used in the RCAC loop, and the missing testing loss and lack of baseline comparisons weaken the evidence for the ANN-based regulation claim. These issues are local and fixable within the manuscript's scope.","major_comments":[{"comment":"The central claim that the closed-loop simulations use the ANN-predicted thrust as the feedback signal yk in Eq. (11) is not explicitly stated in the simulation section. Section III describes both an ideal scenario with true thrust feedback and a practical scenario with ANN-predicted thrust, and Fig. 4 shows the ANN in the loop, but Section IV-A does not identify which signal was used as yk for the RCAC update. Figures 5-8 plot 'commanded and generated thrust,' which could refer to the true model thrust T from Eq. (8) rather than the ANN estimate. The conclusion asserts that the ANN predictions were used as the proxy for thrust, but the simulation section must state this explicitly for each example and, ideally, provide a comparison of ANN-based feedback with direct thrust feedback. Without this, the reader cannot verify the central claim that the combined ANN+RCAC framework regulates thrust using in-situ measurements.","section":"Section IV-A, Eq. (11), and Fig. 4"},{"comment":"The testing loss is explicitly omitted ('not shown for brevity'), but the text states that minimization of both training and testing loss was used to evaluate ANN performance and to select 100 epochs as a compromise against overfitting. Reporting only training loss does not support the claim that the final model generalizes to the closed-loop operating region. The authors should report the testing loss, at least as a table or as an additional line in Fig. 3, and state the final training and testing mean-squared error for the selected architecture.","section":"Section II-B, Fig. 3"},{"comment":"The closed-loop results are reported only as time-history plots; there are no quantitative tracking metrics (e.g., RMS tracking error, settling time, overshoot) and no comparison with a baseline controller or with direct thrust feedback. The error plots on a log scale appear small, but summary statistics are needed to support the statement that the framework 'effectively regulates' thrust, especially for the ramp and harmonic commands where the tracking error may vary over time. A comparison of ANN-based feedback versus true-thrust feedback would directly address the main contribution of the paper.","section":"Section IV-A, Figs. 5-8"}],"minor_comments":[{"comment":"The physical fidelity of the quasi-1D SFRJ model, including the inlet curve fits from Ref. [8] and the regression-rate constants from Ref. [35], is the load-bearing premise for transfer to a real engine. Because the ANN is trained on synthetic data from this same model and the closed-loop plant is the same model, the demonstration is self-consistent but inherits any model bias. This limitation should be stated explicitly in the conclusions, along with an indication of what experimental validation would be needed.","section":"Section II-A"},{"comment":"The symbol r0 is used both for the cowl radius state and for the nominal value in the recurrence r0 = r0 + 0.001uk; using a different symbol for the nominal value would avoid confusion.","section":"Section IV-A, Eq. (19)"},{"comment":"The statement that the sigmoid activation function is applied on all layers 'to return a normalized value of [0,1]' is only strictly correct for the output layer; hidden-layer sigmoid activations do not normalize the network output, and the sentence should be clarified.","section":"Section II-B"},{"comment":"The three panels of Fig. 3 appear to lack axis labels, legends, and clear association with the hyperparameter cases in Table III, making it difficult to interpret the sensitivity study. Please add labels or a caption that maps each panel to a specific hyperparameter sweep.","section":"Fig. 3"},{"comment":"There are several typographical and notation issues: the title header contains 'Regulati on', Section IV reads 'In all subsequence examples', and 'P0 = 10 −5I3' should be typeset as 10^{-5} I_3. These should be corrected in revision.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The paper is within scope for a control-application journal, and the core idea is interesting. The primary issue is the ambiguity about the feedback signal in the closed-loop simulations, which directly affects the validity of the central claim. The missing testing loss and baseline comparisons are secondary but should be addressed. I do not see any evidence of misconduct; the revision should focus on clarity and completeness of the simulation evidence."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Bottom line: this is a solid engineering extension, not a breakthrough. The genuinely new bit is using the variable-geometry inlet as the actuator and feeding an ANN thrust surrogate (inputs: cowl radius, altitude, Pt4, CO mole fraction) into an RCAC adaptive PID for a quasi-1D SFRJ model. Prior RCAC-SFRJ work used direct thrust feedback and heat-flux input, so this is a legitimate increment for the in-situ sensing problem.\n\nWhat it does well: the model is described in enough detail to be reproduced, the training data generation is clear, and the RCAC formulation is standard. The paper reports command following for step, doublet, ramp, and harmonic commands with fixed RCAC hyperparameters, plus a hyperparameter sensitivity sweep. The conclusion explicitly states that simulations used the neural network predictions as the proxy for thrust, which answers the main stress-test worry: the feedback path appears to be the ANN output, not model truth. That said, the main text would be much clearer if Section IV said this directly, and the authors should show a comparison against direct-thrust feedback to prove the surrogate is not silently degrading the loop.\n\nSoft spots, in order of importance. First, the feedback-path ambiguity in Section IV is a real readability flaw, though not a fatal one given the conclusion. Second, there is no baseline: no summary statistics for tracking error, no comparison to a fixed-gain PID or RCAC with true thrust. The claim of robustness to hyperparameter variation is supported only by eyeballing Figure 9. Third, the testing loss is not shown for brevity, a needless omission; one number or a learning curve would remove a lot of doubt about whether the ANN is actually accurate on held-out points. Fourth, the load-bearing assumption is that the quasi-1D model with curve fits from [8] and regression rate constants from [35] represents a real SFRJ. Since both the training data and the plant come from the same model, the framework inherits its fidelity. That is not circularity in the derivation, but it caps the practical claims; the abstract is appropriately careful in saying the SFRJ model.\n\nWho this is for: people working on SFRJ or scramjet thrust control with surrogates, and anyone wanting a worked example of RCAC plus neural-network feedback. A serious referee could give useful guidance, mainly on the missing comparisons and the explicit feedback-path statement. I would not block publication on the science; I would ask for clarifications and a baseline figure.","headline":"A competent engineering extension — ANN + RCAC with variable-inlet actuation for a quasi-1D SFRJ model — publishable with revisions but not a breakthrough.","tokens_in":803,"tokens_out":1685,"would_cite":false,"duration_ms":41555,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["93C40","68T07"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that neural-network thrust estimation from limited in-situ sensors, coupled with a retrospective-cost adaptive controller, can regulate a solid fuel ramjet's thrust in numerical simulation across multiple command types.","keywords":["solid fuel ramjet","thrust estimation","neural network","retrospective cost adaptive control","adaptive PID","in-situ sensing","command following"],"falsifier":"Compare the neural network's thrust estimates against direct load-cell thrust measurements on a ground-test SFRJ with the same geometry and cruise conditions; if the estimates deviate beyond the training tolerance under realistic sensor noise, the framework's in-situ monitoring claim fails.","tokens_in":10525,"feed_emoji":"🚀","tokens_out":6802,"duration_ms":71989,"temperature":0.7,"pith_summary":"This paper tries to show that a solid fuel ramjet (SFRJ) can monitor and regulate its own thrust without a direct thrust sensor, using only measurements that are practical in flight. A neural network is trained on synthetic data from a quasi-one-dimensional SFRJ model to estimate thrust from altitude, combustor pressure, exhaust carbon monoxide fraction, and the variable inlet position. An online adaptive controller, built on retrospective cost adaptive control, then adjusts the inlet geometry to follow thrust commands using those estimates as feedback. Numerical simulations show the same controller tracks step, doublet, ramp, and harmonic commands without retuning. The payoff, if the result transfers beyond the model, is a route to closed-loop SFRJ thrust control with a minimal sensing footprint.","feed_headline":"Neural net and adaptive control regulate ramjet thrust in simulation","feed_subtitle":"Estimates thrust from pressure, CO, altitude, and inlet position; then tunes the controller on the fly.","key_machinery":"The framework's three coupled components are: (1) a quasi-one-dimensional SFRJ plant model whose inlet conditions come from curve fits to axisymmetric RANS simulations and whose fuel regression uses empirical constants, (2) a single-hidden-layer neural network (20 sigmoid neurons) that estimates thrust from altitude, total combustor pressure, exhaust CO fraction, and cowl radius, and (3) an RCAC-based adaptive PID controller whose gains are updated online by minimizing a retrospective cost using recursive least squares. The load-bearing link is that the controller treats the network's estimated thrust as the measured output, so the quality of the closed loop is bounded by the estimator's accuracy on the operating trajectory.","core_discovery":"The central claim is that combining an artificial neural-network thrust estimator with a retrospective-cost adaptive controller yields effective thrust regulation for a solid fuel ramjet with variable-cowl inlet, using limited in-situ measurements instead of direct thrust measurement. The paper demonstrates this on a quasi-one-dimensional feed-forward SFRJ model, with the neural network trained on 125,000 synthetic samples spanning altitude, cowl radius, and port radius. The trained network maps the available on-board measurements to thrust, and the adaptive controller, an RCAC-optimized PID, recursively updates its gains from past tracking error and input data alone. In the reported simulations the closed loop follows step, doublet, ramp, and harmonic thrust commands without changing controller hyperparameters, and the response is insensitive to those hyperparameters over a tested range.","pith_inferences":["The natural next step is to train the network on experimental data (or a validated higher-fidelity model) and test on a hardware-in-the-loop rig; the current evidence is entirely synthetic, so transfer is not yet established.","Because the estimator uses CO, the framework could be extended to estimate and regulate equivalence ratio or combustion efficiency, not just thrust, as long as the corresponding sensors exist.","A robustness gap remains: sensor noise, bias, and time delays in real CO and pressure measurements are not modeled here, so the controller's margins under realistic sensing are an open question.","The paper notes it only sketches the RCAC update proof due to page limits, so a complete stability analysis of the coupled NN-plus-RCAC loop is not provided."],"forward_implications":["If the results hold, SFRJ thrust can be regulated using only sensors that can plausibly fly on a supersonic vehicle, avoiding a direct thrust sensor.","The same controller, with fixed hyperparameters, tracks step, doublet, ramp, and harmonic commands, so no per-mission retuning is needed.","The closed-loop response is robust over a range of RCAC hyperparameters, which helps in practice when plant conditions drift.","The neural network's sensitivity analysis gives a simple guideline: hidden-layer width matters more than activation function or batch size for this thrust-estimation task."],"supporting_citations":[{"why":"Supplies the variable-inlet SFRJ model whose curve fits define the plant and the synthetic training data.","marker":"[8]"},{"why":"Provides the empirical regression-rate constants that set fuel mass flow in the plant model.","marker":"[35]"},{"why":"Defines the RCAC algorithm that updates the adaptive PID gains online.","marker":"[25]"},{"why":"Demonstrates RCAC thrust regulation for an SFRJ with direct thrust measurement, the baseline this work extends to in-situ estimation.","marker":"[11]"},{"why":"Shows CO can be measured in-situ by diode-laser absorption, justifying the sensor choice.","marker":"[41]"},{"why":"Is the deep-learning library used to train the neural network.","marker":"[43]"},{"why":"Is the stochastic optimization method used for network training.","marker":"[44]"}],"fun_headline_variants":["Ramjet thrust tamed by neural net and adaptive control","Neural estimator plus adaptive controller keeps ramjet on target","AI-driven control regulates solid fuel ramjet thrust in simulation","Self-tuning controller and neural net monitor ramjet thrust in situ","Learning framework adjusts ramjet thrust with limited sensor data"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The neural network and the controller are both built and tested on the same quasi-one-dimensional SFRJ model; if that model's inlet curve fits and regression-rate constants do not match a real engine, the thrust estimates and control actions will not transfer.","fun_headline_variants_meta":{"raw":{"variants":["Ramjet thrust tamed by neural net and adaptive control","Neural estimator plus adaptive controller keeps ramjet on target","AI-driven control regulates solid fuel ramjet thrust in simulation","Self-tuning controller and neural net monitor ramjet thrust in situ","Learning framework adjusts ramjet thrust with limited sensor data"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000539,"raw_usage":{"total_tokens":2531,"prompt_tokens":835,"completion_tokens":1696,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":451,"completion_tokens_details":{"reasoning_tokens":1615}},"tokens_in":451,"tokens_out":1696,"duration_ms":14039,"temperature":1.0,"reasoning_tokens":1615,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T05:18:28.425528+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compare the neural network's thrust estimates against direct load-cell thrust measurements on a ground-test SFRJ with the same geometry and cruise conditions; if the estimates deviate beyond the training tolerance under realistic sensor noise, the framework's in-situ monitoring claim fails.","supporting_citations":[{"cited_title":"Augmentation of solid fuel ramjet performance using morphing inlet design,","cited_arxiv_id":null,"evidence_quote":"Supplies the variable-inlet SFRJ model whose curve fits define the plant and the synthetic training data."},{"cited_title":"Investigati on of solid-fuel, dual-mode combustion ramjets,","cited_arxiv_id":null,"evidence_quote":"Provides the empirical regression-rate constants that set fuel mass flow in the plant model."},{"cited_title":"Retrospective cost adaptive control: Pole placement, frequency response , and connections with lqg control,","cited_arxiv_id":null,"evidence_quote":"Defines the RCAC algorithm that updates the adaptive PID gains online."},{"cited_title":"Learning-based adaptive thrust regulation of solid fuel ramjet,","cited_arxiv_id":null,"evidence_quote":"Demonstrates RCAC thrust regulation for an SFRJ with direct thrust measurement, the baseline this work extends to in-situ estimation."},{"cited_title":"In situ combustion measurements of co with diode-laser absorption near 2.3 µm,","cited_arxiv_id":null,"evidence_quote":"Shows CO can be measured in-situ by diode-laser absorption, justifying the sensor choice."}],"review_version":1}