{"id":"c3219f28-1465-4ea6-a93d-0fffbd49af39","arxiv_id":"2502.00532","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"A tiny neural network that imitates a hand-edited PI signal reduced overshoot in simulated FOC motor control by up to 87.5%, but its inference time misses the real-time microcontroller budget.","lead":"A 1,400-parameter neural network was trained to correct overshoot in the speed controller of an electric motor and was shrunk for a microcontroller. In two custom simulation signals it cut overshoot by up to 87.5%, but its measured inference time was 4 to 11 times slower than the 30 kHz control loop requires.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"All reported MCU inference times exceed the 33.33 µs control period, so the claimed MCU deployment fails by the paper's own measurements and the simulated overshoot gains assume zero-latency inference.","rationale":"The central claim has two parts: (i) TinyFC reduces overshoot in FOC simulation, and (ii) it is optimized for microcontroller deployment. My stress-test targets part (ii), because it is falsified by the paper's own Table 5 and Section 5. The reader's weakest-assumption concerns the hand-constructed training target in Eqs. (1)–(3), which is a real and serious issue, but it requires a judgment about what counts as an ideal current reference. The timing contradiction is internal, purely arithmetic, and does not depend on interpreting the training objective: every measured inference time exceeds the control period that the paper itself sets. This makes the deployment claim unsupported even if the training target were accepted, and it also undermines the transferability of the closed-loop simulation results. I find no reason to dispute the reader's REJECT verdict; my concern reinforces it, so the verdict is unchanged. The paper does contain useful workflow elements and explicit timing measurements, but those measurements contradict the headline claim rather than support it.","tokens_in":9133,"tokens_out":5210,"duration_ms":53399,"concrete_test":"Modify the Simulink FOC model to delay the TinyFC output Δi_q by the measured inference times (207.4 µs for TinyFC, 232.2 µs for pruned, 371.7 µs for quantized pruned) before applying Eq. (4), then rerun test cases 1 and 2 and compare max overshoot and deviations against Tables 3 and 4. If overshoot reduction disappears or the loop becomes unstable, the deployment claim is refuted. If performance is preserved, repeat with true hardware-in-the-loop on the NUCLEO-G474RE to confirm the timing model.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 5 states that the 30 kHz PWM requires inputs every 33.33 µs and warns that exceeding this time causes system delays and performance issues. Table 5 reports every inference time on the NUCLEO-G474RE as above that bound: TinyFC 207.4 µs, pruned 232.2 µs, quantized pruned 371.7 µs, and HPO 127.6/144.8 µs. Even the fastest variant is about 4x the control period, and the pruned model that 'achieves complete overshoot elimination' in test case 1 runs at 232.2 µs, about 7x the bound. The closed-loop results in Tables 3 and 4 are obtained in Simulink without modeling this measured latency, effectively assuming that Δi_q is available instantaneously. A real MCU inserting a 207–372 µs delay into a 33.33 µs control loop would alter the loop dynamics, potentially removing the reported overshoot reduction or causing instability. The paper's own conclusion concedes that 'optimizing TinyFC inference for MCU timing remains a challenge.' Therefore the central claim of a microcontroller-optimized network that improves FOC is not supported by the evidence as presented.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes TinyFC, a 1,400-parameter feedforward neural network inserted into the speed loop of PI-based field-oriented control for a PMSM. The network takes reference speed, measured speed, and the PI quadrature-current reference as inputs and outputs a compensation value Δi_q that is added to the PI reference. Training data come from Simulink simulations of two custom speed profiles; the target Δi_q^GT is defined as the difference between the PI current reference and a manually adjusted version obtained by saturation (Eq. 1) or exponential rectification (Eq. 2). The authors compare the baseline, HPO, and pruned variants, all quantized, in terms of MSE, speed deviation and overshoot, and deployed inference time on a NUCLEO-G474RE. They report up to 87.5% overshoot reduction with TinyFC and about 100% reduction with the pruned model in test case 1, but they also report inference times from 127.6 to 371.7 µs, all exceeding the 33.33 µs period of the 30 kHz control loop.","tokens_in":9395,"tokens_out":5789,"duration_ms":58859,"significance":"If the central claims were supported, the paper would be a relevant demonstration of a very small neural compensator for motor control, with strengths including actual MCU inference measurements through ST Edge AI, a comparison of several compression techniques, and a public code/data repository. The paper is also honest in acknowledging that MCU timing remains a challenge. However, two load-bearing issues undermine the main claims: the training target is constructed from the PI controller's own current signal by manual overshoot removal, making the measured overshoot reduction partly a definitional artifact; and the measured inference times exceed the control period by factors of roughly 4 to 11, so the closed-loop simulation results in Tables 3 and 4 implicitly assume zero-latency compensation. These issues affect the core message of a microcontroller-optimized network that improves FOC, so the manuscript as it stands does not support its conclusions.","major_comments":[{"comment":"The network is trained to predict Δi_q^GT = x(t) − x_adj(t), where x_adj(t) is the PI reference current after manual overshoot removal with hand-chosen parameters C and τ. Adding this prediction to the PI reference via Eq. (4) means that, by construction, the loop tracks the hand-edited reference whenever the network fits its target. The overshoot reductions in Tables 3 and 4 therefore do not validate that the network improves motor control; they mainly show that the network can reproduce a manually edited current signal. An independent validation target based on physical control criteria, such as closed-loop stability, actuator limits, or disturbance rejection, is needed before the result can be attributed to the neural compensator.","section":"Section 4, Eqs. (1)-(3)"},{"comment":"The paper correctly states that the 30 kHz PWM requires inputs every 33.33 µs and that exceeding this time causes delays and performance issues. Yet every inference time reported in Table 5 exceeds this bound: TinyFC 207.4 µs, HPO 144.8 and 127.6 µs, pruned 232.2 and 215.2 µs, and quantized-pruned 371.7 and 361.4 µs. The Simulink results in Tables 3 and 4 do not model this measured latency, so the reported overshoot improvements assume the compensation is applied instantaneously. This directly contradicts the abstract's claim that the model fits the computational constraints of a microcontroller, and it is consistent with the paper's own concluding admission that 'optimizing TinyFC inference for MCU timing remains a challenge.'","section":"Section 5, first paragraph, and Section 6.3, Table 5"},{"comment":"The HPO model, one of the three proposed optimization outcomes, performs worse than the PI baseline in test case 2: Table 4 reports a maximum overshoot of 1.24 for HPO, versus 0.25 for PI in Table 3. The paper states that 'the HPO model did not perform satisfactorily within the FOC loop during this study (this case is not shown in this paper)' and then proceeds without further analysis. As a result, the claim that hyperparameter optimization is a successful route in the proposed workflow is unsupported, and the trade-off discussion among the variants is incomplete.","section":"Section 6.1 and Table 4"},{"comment":"Equation (1) defines x_adj(t) = min{|x(t)|, C}, which is always nonnegative. If x(t) denotes the quadrature-current reference, this definition cannot represent a valid reference during deceleration or negative-torque operation, because the sign information is discarded. Even if the test signals used here happen to avoid negative values, the equation as written is not a general saturation operator and should be replaced by a sign-preserving saturation or explicitly restricted to the positive-current case.","section":"Section 4, Eq. (1)"}],"minor_comments":[{"comment":"The notation x(t) is used inconsistently: it sometimes denotes the measured speed, sometimes the quadrature current, and sometimes the 'response signal.' Please define each symbol explicitly at first use.","section":"Section 4, Eqs. (1)-(3)"},{"comment":"The test cases are described only qualitatively as '2 transitions per second' and '10 speed transitions per second.' The exact reference-speed sequences should be given numerically or as a data file so that the experiments are reproducible.","section":"Section 3.2"},{"comment":"The quantization description omits implementation details such as per-tensor versus per-channel quantization, the calibration dataset, and the integer runtime used by ST Edge AI. These details are necessary to reproduce the reported memory and timing numbers.","section":"Section 5.3"},{"comment":"The table lists 'Act' and 'Lib' memory components, but the meaning of these entries and their units are not explained in the text; please clarify.","section":"Section 6.3, Table 5"}],"recommendation":"reject","confidential_remarks":"The reader's report aligns with my own reading: the inference-time measurements alone contradict the central real-time feasibility claim, and the training-target construction makes the overshoot result largely self-fulfilling. I do not see a revision that could fix both within the current scope without a substantially new validation methodology and a new deployment result."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper's own timing measurements contradict its headline claim. Every inference time on the NUCLEO-G474RE (127.6–371.7 µs) exceeds the 33.33 µs control period, and the closed-loop overshoot reductions are computed in Simulink without modeling that latency. The stress-test note is correct: a real MCU inserting a 200+ µs delay into a 30 kHz loop would change the dynamics, and the reported gains may not survive. The paper even concedes in its conclusion that 'optimizing TinyFC inference for MCU timing remains a challenge,' so this is not a hidden flaw—but the abstract overstates what is demonstrated.\n\nWhat is genuinely useful here is the workflow and the measured deployment data. The authors ship code and data, report MACC, flash, RAM, and inference times for TinyFC, HPO, pruned, and quantized variants, and show that pruning can sometimes make inference slower—a non-obvious, practical result. The idea of training a tiny feed-forward network on the residual between a PI current command and a hand-repaired version of that command is a legitimate control-augmentation scheme, and the two test cases are clearly described. If the paper were reframed as a cautionary study of TinyML deployment in FOC, or if the authors fixed the latency by restructuring the network or the control period, it would have value.\n\nThe soft spots are proportional. The training target is the PI output after saturation with hand-chosen C (Eq. 1) or exponential rectification with hand-chosen tau (Eq. 2); the overshoot reduction on the same test cases is partly an artifact of the target construction. Parameters C and tau are not tuned systematically, and the evaluation is in-sample with no noise or disturbance. These are serious but not fatal to the core idea; they mean the measured improvement should not be taken at face value. The more load-bearing problem is the timing, because it invalidates the deployment claim entirely.\n\nFor you, I'd read this if you work on TinyML control loops or on latency-aware neural network deployment. It is not a rejection-worthy paper in my view: a serious referee could push the authors to either redesign the network to fit the 33.33 µs budget or present the work as a latency-measurement study. I'd accept it for peer review with the expectation of major revision.","headline":"The paper's own timing table sinks its central MCU claim, but the reproducible workflow and measured latency numbers are worth a second look if reframed.","tokens_in":9935,"tokens_out":1424,"would_cite":false,"duration_ms":16026,"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 1,400-parameter feed-forward network, TinyFC, added to a PI speed loop learns to correct the quadrature-current reference, cutting overshoot by up to 87.5% in simulation while keeping the model deployable on a micro-controller.","keywords":["tiny neural networks","field-oriented control","PMSM","PI controller","overshoot reduction","microcontroller deployment","network pruning","quantization"],"falsifier":"Run the two speed profiles on a physical PMSM drive with the same motor and the NUCLEO-G474RE board, with the pruned and quantized TinyFC inserted in the loop; if the 87.5% or complete overshoot reduction does not reproduce, or the motor current becomes unstable, the central claim fails.","tokens_in":8934,"feed_emoji":"⚡","tokens_out":9759,"duration_ms":86134,"temperature":0.7,"pith_summary":"This paper tries to establish that a tiny feed-forward network, TinyFC, can act as a corrective add-on to a classical PI controller in field-oriented motor control. In simulation, the 1,400-parameter network learns to repair the PI controller's quadrature-current reference, cutting maximum speed overshoot by up to 87.5% in one test case and 68% in another, with a pruned version eliminating overshoot entirely in the first case. The appeal of the claim is that a network small enough to fit a micro-controller can fix a known weakness of simple PI control without the latency and complexity of heavier alternatives such as model-predictive control. The paper also reports that quantizing the pruned model to 8-bit integers keeps its weights near 1 KiB, the scale needed for embedded deployment.","feed_headline":"A 1,400-parameter network cuts motor overshoot up to 87.5 percent","feed_subtitle":"A tiny feed-forward compensator learns to correct a PI controller's current reference and fits on a micro-controller.","key_machinery":"The load-bearing object is TinyFC, a feed-forward network with two branches of five fully connected layers plus residual connections, whose output is a current compensation $\\Delta i_q$ constrained to $[-1,1]$ by a $\\tanh$ activation and added to the PI controller's quadrature-current reference. Its training signal, defined by Eq. (3), is the difference between the raw PI current and a hand-adjusted ideal version: saturation at a chosen threshold $C$ for test case 1, exponential rectification with a chosen time constant $\\tau$ for test case 2. The network learns this correction from a 300,001-sample-per-test-case dataset generated by the Simulink FOC model. This target-construction step is what makes the network learn to remove overshoot, and it is the fragile part of the pipeline.","core_discovery":"The central claim is that a small fully connected network, TinyFC, can be trained to output the quadrature-current correction $\\Delta i_q$ that compensates for what a PI speed controller gets wrong, and that adding this correction to the PI reference current inside a field-oriented PMSM drive reduces overshoot and tracking error. The network takes reference speed, measured speed, and the PI-predicted quadrature current as inputs. Its training target is constructed by taking the PI's own quadrature-current signal and removing the sections the authors identify as overshoots, either by saturating at a threshold $C$ or by exponentially rectifying each interval with time constant $\\tau$; the network learns the difference between the raw and the adjusted signals. In the two simulation test cases, the augmented loop reduces maximum overshoot by 87.5% and 68%, and the PCA-pruned version eliminates overshoot completely in the first case.","pith_inferences":["If the hand-edited target is replaced by a signal from an established reference controller, the same pipeline could produce a TinyFC teacher without hand-chosen thresholds or time constants.","A loss function that penalizes overshoot and current excursions directly might fix the failure of MSE-based model selection; the paper points toward physics-informed networks but does not test them.","The reported MCU inference times (127.6 to 371.7 microseconds) are above the 33.33 microsecond PWM period, so whether the optimized models can actually run inside the real-time loop remains an open hardware-in-the-loop question."],"forward_implications":["In the simulated drive, adding TinyFC cuts maximum overshoot by 87.5% in test case 1 and 68% in test case 2, while average deviation falls by 60% and 16.7%.","The PCA-pruned TinyFC eliminates overshoot completely in test case 1 and lowers it to 0.03 in test case 2, with only a small increase in maximum deviation.","Quantization to 8-bit integers brings the pruned model's weight storage down to 1.37 KiB in case 1 and 1.07 KiB in case 2, within the flash budget of the NUCLEO-G474RE board.","Because the network only adds a correction to the PI reference, the PI controller remains in place; the paper explicitly describes TinyFC as supportive rather than a replacement.","MSE during training does not predict loop performance: the hyperparameter-optimized model with fewer parameters and low MSE raised overshoot by up to 100% in case 2, so control-loop metrics must be part of model selection."],"supporting_citations":[{"why":"Supplies the FOC dataset and simulation workflow used to generate the training sequences for both test cases.","marker":"[16]"},{"why":"Provides the Simulink model of field-oriented control for a 3-phase PMSM that serves as the virtual testbed.","marker":"[17]"},{"why":"Defines the motor and inverter hardware parameters used in the simulation.","marker":"[23]"},{"why":"Defines the target NUCLEO-G474RE board and its flash/RAM limits used for deployability evaluation.","marker":"[24]"},{"why":"Tool used for post-training quantization and for profiling the models on real MCUs.","marker":"[25]"},{"why":"Basis for the test signals and the deviation and overshoot metrics used to compare PI and TinyFC control.","marker":"[18]"},{"why":"Cited for the threshold/saturation operation used to construct the adjusted quadrature-current target in test case 1.","marker":"[3]"},{"why":"Provides the PCA-based projection pruning method used to produce the pruned models that eliminate overshoot.","marker":"[28]"}],"fun_headline_variants":["Tiny neural net removes motor overshoot in field-oriented control","Microcontroller neural net slashes motor overshoot by 87.5%","Pruned tiny network kills motor overshoot completely","Field-oriented control gets a tiny neural boost for MCUs","Nail motor control with a 1400-param neural net"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the hand-edited quadrature-current signal, saturated at a chosen threshold $C$ or exponentially rectified with a chosen time constant $\\tau$, is the ideal response the motor should follow; if that edited signal is not physically sound, the measured overshoot reduction is an artifact of the target rather than evidence that TinyFC improves control.","fun_headline_variants_meta":{"raw":{"variants":["Tiny neural net removes motor overshoot in field-oriented control","Microcontroller neural net slashes motor overshoot by 87.5%","Pruned tiny network kills motor overshoot completely","Field-oriented control gets a tiny neural boost for MCUs","Nail motor control with a 1400-param neural net"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000618,"raw_usage":{"total_tokens":2854,"prompt_tokens":917,"completion_tokens":1937,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":533,"completion_tokens_details":{"reasoning_tokens":1861}},"tokens_in":533,"tokens_out":1937,"duration_ms":13694,"temperature":1.0,"reasoning_tokens":1861,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T18:39:11.989901+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the two speed profiles on a physical PMSM drive with the same motor and the NUCLEO-G474RE board, with the pruned and quantized TinyFC inserted in the loop; if the 87.5% or complete overshoot reduction does not reproduce, or the motor current becomes unstable, the central claim fails.","supporting_citations":[{"cited_title":"Field oriented control dataset of a 3-phase permanent magnet synchronous motor.Data in Brief, 47:109002, 2023","cited_arxiv_id":null,"evidence_quote":"Supplies the FOC dataset and simulation workflow used to generate the training sequences for both test cases."},{"cited_title":"Modelling the field oriented control applied to a 3-phase permanent magnet synchronous motor","cited_arxiv_id":null,"evidence_quote":"Provides the Simulink model of field-oriented control for a 3-phase PMSM that serves as the virtual testbed."},{"cited_title":"Motor control nucleo pack with nucleo-f302r8 and x-nucleo- ihm07m1","cited_arxiv_id":null,"evidence_quote":"Defines the motor and inverter hardware parameters used in the simulation."},{"cited_title":"Stm32 nucleo-64 development board with stm32g474re mcu","cited_arxiv_id":null,"evidence_quote":"Defines the target NUCLEO-G474RE board and its flash/RAM limits used for deployability evaluation."},{"cited_title":"St edge ai developer cloud","cited_arxiv_id":null,"evidence_quote":"Tool used for post-training quantization and for profiling the models on real MCUs."},{"cited_title":"Discrete-Time Control Systems","cited_arxiv_id":null,"evidence_quote":"Basis for the test signals and the deviation and overshoot metrics used to compare PI and TinyFC control."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Cited for the threshold/saturation operation used to construct the adjusted quadrature-current target in test case 1."},{"cited_title":"Linear pruning tech- niques for neural networks-based on projection latent structure","cited_arxiv_id":null,"evidence_quote":"Provides the PCA-based projection pruning method used to produce the pruned models that eliminate overshoot."}],"review_version":1}