{"id":"5700b7bf-85f5-41b1-aadf-1e3f510c6409","arxiv_id":"2502.04417","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A 2.4 MB neural network surrogate reproduces EPA MOVES CO2 emissions with 6.013% MAPE on millions of test scenarios, enabling fast microscopic emission modeling.","lead":"The authors reverse-engineer EPA's MOVES vehicle emission simulator into a 109 million point dataset, then train a 2.4 MB neural network that reproduces MOVES CO2 estimates within about 6% average error. This lightweight surrogate makes microscopic, real-time emission estimation feasible for eco-driving and traffic control research.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (3)'s finite-difference labels may not correspond to MOVES's instantaneous (v,a) emission; the appended second is evaluated at v+a and also perturbs the preceding second, so the 6.013% MAPE may be against a shifted target.","rationale":"The paper has real strengths: a large released dataset, an extensive validation set, and a concrete eco-driving use case. The 6.013% MAPE over more than two million cycle-level comparisons is meaningful evidence that the surrogate captures something close to MOVES totals. However, cycle-total comparisons are relatively insensitive to the kind of systematic relabeling identified above: a one-step shift in the speed/acceleration state can partially cancel when summed over a trajectory, especially with approximately symmetric acceleration distributions. The central claim is specifically about microscopic, instantaneous emissions and differentiability for control; that claim inherits the correctness of Eq. (3). Since the paper's own Eq. (4) contains an off-by-one inconsistency for the same cycle, independent verification of the label construction is essential. This supports the reader's conditional verdict rather than a rejection: if the probe confirms the labels, the reverse-engineering pipeline is sound; if it does not, the surrogate can be retrained on corrected labels without abandoning the overall methodology. The concern is about the argument, not the authors; the released code makes the proposed check straightforward.","tokens_in":12752,"tokens_out":13200,"duration_ms":152713,"concrete_test":"Use the released MOVESRE generation scripts to extract the label for one fixed (v=15 m/s, a=1 m/s2) with the paper's n=5 cycle, and also by a direct probe cycle whose target second has speed 15 and acceleration 1 (e.g., speeds [14,15,16,...] with a constant-speed baseline). Compare the two deltas, checking also whether Eq. (4)'s mode fractions in the code are n/(n+1) or (n-1)/n. If the deltas differ beyond MOVES's numerical tolerance, Eq. (3) does not label e(15,1) and the surrogate's instantaneous semantics need revision before the headline error rate is meaningful.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"Load-bearing concern: the reverse-engineering step that defines the training labels. In Eq. (3), e(v,a,X) is computed as E(tau',X)-E(tau,X), where tau is n seconds at speed v and tau' is those n seconds plus an appended step v'=v+a (Eq. (2)). This subtraction does not isolate a single second at state (v,a). Under MOVES's finite-difference speed/acceleration convention, the appended step changes the acceleration of the last baseline second as well as introducing the new second, and the new second has speed v+a, not v. So the extracted label is either a two-second mixture or a label for (v+a,a); in neither case is it MOVES's instantaneous emission at (v,a). The off-by-one in Eq. (4) — p1=(n-1)/n, p2=1/n for an (n+1)-step cycle — is a concrete symptom of this unresolved bookkeeping. If the labels are systematically shifted, the surrogate is trained to reproduce a different microscopic target, and the reported validation MAPE on summed cycle emissions cannot certify the claimed microscopic fidelity.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces NeuralMOVES, a lightweight neural network surrogate for the EPA MOVES CO2 emission simulator. The authors first reverse-engineer MOVES by constructing pairs of driving cycles and differencing their total emissions to obtain second-by-second emission labels e(v,a,X), yielding a 9.89 GB dataset MOVESRE with over 109 million entries. They then train a two-layer, five-neuron neural network on this dataset, obtaining a reported 6.013% MAPE against MOVES over 2.3 million validation scenarios that combine 100 synthetic driving cycles with various environmental and vehicle settings. The model is 2.4 MB, differentiable, and runs in milliseconds. The paper also demonstrates an MPC-based eco-driving use case and publicly releases the code and dataset.","tokens_in":12998,"tokens_out":4842,"duration_ms":52583,"significance":"If the reverse-engineering method is correct, this is a practically valuable contribution: it would make MOVES-compatible microscopic CO2 emission estimates available in a lightweight, differentiable, real-time form, with open-source code and a large public dataset. The paper explicitly names its assumptions and limitations, including the unexplained -5% grade surge, which is commendable. However, the central label-construction step (Eq. (3)-(4)) contains an apparent off-by-one error and an unresolved ambiguity about which MOVES operating-mode second is isolated by the differencing procedure. Since all training labels and all validation claims rest on this step, the current manuscript does not yet establish the claimed microscopic fidelity.","major_comments":[{"comment":"The mode fractions for the custom cycle τ' are off by one. The cycle τ' is defined in Eq. (2) as an (n+1)-step cycle: n steps at speed v plus one appended step at v' = v + a∆t. The fractions of time spent in the two relevant speed/acceleration states should therefore be n/(n+1) and 1/(n+1), not (n-1)/n and 1/n as written. The authors must correct this, or explicitly justify why (n-1)/n is the correct fraction; as written, Eq. (4) is inconsistent with Eq. (2) and undermines the quantitative basis of the reversed-engineered emissions.","section":"§2.1.1, Eq. (4)"},{"comment":"The central extraction equation e(v,a,X) = E(τ',X) - E(τ,X) does not necessarily isolate the instantaneous emission of a single second at state (v,a). In MOVES's operating-mode framework, the appended step at speed v+a also changes the acceleration of the immediately preceding second (if acceleration is computed from the following speed), or the appended second itself is in state (v+a,a), not (v,a). The paper does not specify MOVES's exact speed/acceleration binning convention, so the extracted label e(v,a,X) may correspond to a two-second mixture or to a different (v,a) point. Because every training example and every subsequent validation depends on this label definition, the authors must provide an explicit mapping from cycle second-states to MOVES operating modes and demonstrate which seconds are included in the difference, with a concrete example for a small n (e.g., n=2).","section":"§2.1.1, Eq. (3)"},{"comment":"The acknowledged emission surge at a road grade of -5%, whose cause remains unclear, is a red flag for the quality of the reverse-engineered dataset. If the surge is an artifact of the differencing procedure (for instance, a symptom of the off-by-one or mode-mixing issue in Eq. (3)), the surrogate will encode this artifact. The authors should either identify the physical or software cause of the surge or demonstrate through targeted experiments that the validation MAPE of 6.013% is insensitive to this anomaly.","section":"§2.1.3"},{"comment":"The validation set varies driving dynamics broadly but uses environmental and vehicle factors on (or very near) the same discrete grid used to generate the training data (Table 1). The reported 6.013% MAPE therefore primarily measures interpolation fidelity over the static factors, not generalization to unseen environmental or vehicle conditions. To substantiate the claim of accuracy across 'diverse scenarios', the authors should report results on held-out static-factor values (e.g., continuous temperatures/humidities or grades not used in training) or explicitly split the MOVESRE grid into training and validation sets.","section":"§3"}],"minor_comments":[{"comment":"There is a numerical inconsistency: the text states '22,869×100=2,296,900' and then refers to 'all 2,286,900 evaluations'. The correct product is 2,286,900; please fix the typo.","section":"§3"},{"comment":"The MAPE definition divides by e_i, which is undefined if e_i = 0. The paper should specify how zero or near-zero emission values are handled, especially since the model is later clipped at the idling emission value.","section":"§2.2, Eq. (9)"},{"comment":"The abstract uses 'Mean Average Percentage Error', which should be 'Mean Absolute Percentage Error (MAPE)' to match the body of the paper.","section":"Abstract"},{"comment":"The table reports an '11 epochs' neural network with MAPE 149.73%, which is far worse than the 300-epoch models; it would be helpful to state whether this row is an ablation for early stopping and why it is included in the main architecture comparison.","section":"Table 2"},{"comment":"The notation 'max{eNN, eidling} 7→ e' is ambiguous; it should be written as a definition, e.g., e = max(eNN, eidling), to clarify that the output is the truncation, not an additional mapping.","section":"§2.2, Eq. (10)"}],"recommendation":"major_revision","confidential_remarks":"The paper is likely to be of interest to the transportation modeling community, and the open-source release plus the large dataset are valuable. However, the referee has serious concerns about the correctness of the reverse-engineering step that defines the target labels. The off-by-one in Eq. (4) and the unspecified MOVES speed/acceleration binning convention make it impossible to verify that e(v,a,X) is truly a single-second emission. I would urge the editor to require the authors to provide a detailed example of the extraction for a small cycle, ideally with a step-by-step mapping to MOVES operating modes, and to make the extraction code part of the repository. The validation should also be strengthened with out-of-sample static factors. These are fixable within the scope of the paper, so a major-revision recommendation seems appropriate rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Worth a look for the dataset and the practicality, but the reverse-engineering math has a real bookkeeping problem that undermines the claim to have extracted MOVES's true instantaneous emissions.\n\nWhat's genuinely new: the MOVESRE dataset — 109M extracted (v,a) points across grade, temperature, humidity, vehicle type, age, fuel — and the difference-of-cycles idea for prying microscopic output out of a macroscopic simulator. The resulting 2.4 MB two-layer MLP is a nice engineering result: 6.013% MAPE against MOVES on 2.3M cycle-level evaluations, differentiable, millisecond inference. The eco-driving MPC demo is a good illustration. The authors are also transparent about the -5% grade anomaly and the 198k corrupted points.\n\nNow the soft spots. The central extraction, Eq. (3), subtracts a constant-speed n-second cycle from the same cycle with one appended second at v+a. As the stress test notes, that does not isolate a second at (v,a). The appended second is at speed v+a, and it changes the acceleration of the previous second from 0 to a. The difference is therefore a mixture — roughly e(v,a) + e(v+a,0) - e(v,0) — not e(v,a). The off-by-one in Eq. (4) (p1=(n-1)/n, p2=1/n for an n+1-second cycle) is a symptom of this muddled bookkeeping. If the labels are a shifted target, the surrogate is trained on something other than MOVES instantaneous emissions. The cycle-level validation can't resolve this: summing a surrogate trained on a wrong microscopic target can still reproduce cycle totals on the tested cycles because the errors correlate and cancel. The validation static factors are on the same grid as training, so the 6% MAPE is essentially an interpolation check for those factors, not a generalization claim. Also, the paper says the 9.89 GB dataset is publicly available, but the only link is the surrogate repo — no dataset link.\n\nNone of this kills the practical value: if you want a fast, differentiable estimator of MOVES cycle totals, NeuralMOVES looks like it works at the 6% level on the tested distributions. But the scientific claim — that this is a reverse-engineered microscopic MOVES — needs a corrected derivation and a direct check that the extracted labels are consistent (e.g., by showing that the extracted field, integrated over arbitrary cycles, reproduces MOVES totals; ideally also by comparing to MOVES-Matrix or a direct one-second MOVES run if feasible). The -5% grade surge should be investigated rather than just reported.\n\nThis is a paper for transportation/eco-driving researchers who need a usable MOVES surrogate. It deserves a serious referee, but the referee should send it back for a major revision on the extraction math and the validation claims.","headline":"A useful practical surrogate for MOVES, but the reverse-engineering step has a bookkeeping flaw that calls the microscopic-label claim into question.","tokens_in":13523,"tokens_out":5766,"would_cite":true,"duration_ms":54478,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"By reverse engineering the MOVES simulator into a 109-million-point dataset and compressing it into a 2.4 MB neural network, NeuralMOVES reproduces MOVES CO2 estimates with a 6.013% mean absolute percentage error while staying…","keywords":["MOVES","vehicle emissions","surrogate model","reverse engineering","neural network","CO2 estimation","eco-driving","microscopic modeling"],"falsifier":"Construct a short driving cycle whose total emissions are compared against the sum of NeuralMOVES's second-by-second predictions over the same trace; if the summed prediction differs from MOVES's cycle total by more than the reported 6% for cycles that toggle rapidly between operating modes, the additive decomposition underlying the reverse engineering is inadequate. A stronger test is to run the same speed-acceleration trace in both MOVES and NeuralMOVES and compare against a chassis dynamometer measurement of a real vehicle, which would reveal whether the surrogate inherits MOVES's biases without its safeguards.","tokens_in":12572,"feed_emoji":"🚗","tokens_out":7742,"duration_ms":71037,"temperature":0.7,"pith_summary":"NeuralMOVES sets out to make MOVES, a certified vehicle emission simulator, usable in microscopic, real-time applications. The authors reverse engineer MOVES by running it on paired driving cycles that differ by a single appended acceleration step, subtracting the baseline cycle's total emissions to isolate the instantaneous emission of that step. This produces a 9.89 GB dataset of over 109 million speed-acceleration-environment-vehicle combinations, which is then compressed into a 2.4 MB neural network with two hidden layers of five neurons. Across over two million test scenarios built from five driving-behavior generators, the surrogate matches MOVES total emissions to a 6.013% mean absolute percentage error, and its differentiability and millisecond evaluation make it suitable for model predictive control, as shown in a signalized-intersection eco-driving case.","feed_headline":"2.4 MB neural net matches MOVES CO2 estimates to 6% error","feed_subtitle":"The lightweight, differentiable model makes MOVES-compatible emissions practical for real-time eco-driving control.","key_machinery":"The central object is the reverse-engineering identity $e(v, a, X) = E(\\tau', X) - E(\\tau, X)$, which treats MOVES as a black box and isolates the instantaneous emission of an acceleration step by subtracting the total emissions of a constant-speed baseline cycle from those of the same cycle with one appended step. MOVES total emissions are modeled as a mode-fraction-weighted sum of per-mode emission rates over a one-hour operating window, scaled linearly to cycle length, so the subtraction is only valid if emissions are additive over time and operating modes. The surrogate itself is a tanh-activated neural network with two hidden layers of five units, trained to minimize mean absolute percentage error; its differentiability is what allows it to be embedded in gradient-based optimization.","core_discovery":"The paper claims that an industrial-grade macroscopic emission simulator can be reverse-engineered into a compact, differentiable, microscopic surrogate that stays faithful to the original. Concretely, NeuralMOVES maps speed, acceleration, road grade, temperature, humidity, vehicle type, age, and fuel type to an instantaneous CO2 emission rate, with outputs truncated below the idling emission, and the resulting model reproduces MOVES's total cycle emissions within roughly 6% mean absolute percentage error across diverse trajectories. This is achieved with a 2.4 MB architecture, a 4800x reduction over the extracted dataset, enabling emissions to be computed on the fly inside a control loop.","pith_inferences":["If the slight positive skew in the error distribution is systematic, control policies optimized with NeuralMOVES could systematically under-value aggressive acceleration; a sensitivity analysis replacing the surrogate with the 10 GB lookup table inside the same MPC loop would reveal whether optimal trajectories change.","The apparent off-by-one in the mode fractions suggests the extracted instantaneous emissions may be least accurate at high acceleration values where the appended step occupies a large share of the cycle; varying the baseline length n and checking whether NeuralMOVES accuracy degrades would test this directly.","Reverse engineering running emissions for CO2 likely extends to other pollutants with separated running emission rates, so the same pipeline could yield a multi-pollutant microscopic surrogate."],"forward_implications":["Researchers can perform MOVES-compatible emissions analysis without installing MOVES, since the 2.4 MB surrogate is open-sourced and runs on continuous input ranges.","Gradient-based eco-driving controllers can directly minimize emission predictions, because NeuralMOVES is differentiable and evaluates in milliseconds.","The reverse-engineering framework offers a template to distill other industrial-grade simulation software into fast surrogates, subject to legal permission.","Because the model decouples emissions from a fixed U.S. location, users can supply their own temperature, humidity, and road grade to estimate emissions for non-U.S. or customized scenarios."],"supporting_citations":[{"why":"the MOVES simulator being reverse-engineered; supplies the ground-truth emission outputs the surrogate must match.","marker":"(U.S. Environmental Protection Agency, 2022)"},{"why":"MOVES-Matrix, the prior lookup-table approach that NeuralMOVES contrasts with by removing complex inputs and 100 GB storage needs.","marker":"(Liu et al., 2016)"},{"why":"MOVEStar, the prior lightweight MOVES replica that lacks accuracy measures and supports only two vehicle types.","marker":"(Wang et al., 2020)"},{"why":"VT-CPFM, an alternative power-based microscopic model used to show emission profiles and optimal eco-driving trajectories differ from MOVES-based ones.","marker":"(Park et al., 2013)"},{"why":"supplies the reinforcement-learning eco-driving policy used both to generate validation cycles and as the control use case.","marker":"(Jayawardana and Wu, 2022)"},{"why":"IDM, the car-following model that generates intersection-approach validation cycles.","marker":"(Treiber et al., 2000)"}],"fun_headline_variants":["6% error: 2.4 MB neural net replaces MOVES for real-time CO2","Reverse-engineered MOVES into a 2.4 MB neural net with 6% error","Microscopic CO2 model: 2.4 MB, 6% error, real-time ready","From macro to micro: NeuralMOVES shrinks MOVES to 2.4 MB","NeuralMOVES: 2.4 MB surrogate for MOVES, 6% error, real-time"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reverse-engineering step assumes that MOVES total emissions decompose additively over time steps and operating modes, so that the difference between two driving cycles equals the emission of the single newly added step; the paper's stated mode fractions for the appended cycle contain an off-by-one inconsistency that would bias the extracted instantaneous emissions if MOVES treats mode transitions nonlinearly.","fun_headline_variants_meta":{"raw":{"variants":["6% error: 2.4 MB neural net replaces MOVES for real-time CO2","Reverse-engineered MOVES into a 2.4 MB neural net with 6% error","Microscopic CO2 model: 2.4 MB, 6% error, real-time ready","From macro to micro: NeuralMOVES shrinks MOVES to 2.4 MB","NeuralMOVES: 2.4 MB surrogate for MOVES, 6% error, real-time"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000792,"raw_usage":{"total_tokens":3476,"prompt_tokens":919,"completion_tokens":2557,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":535,"completion_tokens_details":{"reasoning_tokens":2432}},"tokens_in":535,"tokens_out":2557,"duration_ms":18583,"temperature":1.0,"reasoning_tokens":2432,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T23:29:00.705749+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Construct a short driving cycle whose total emissions are compared against the sum of NeuralMOVES's second-by-second predictions over the same trace; if the summed prediction differs from MOVES's cycle total by more than the reported 6% for cycles that toggle rapidly between operating modes, the additive decomposition underlying the reverse engineering is inadequate. A stronger test is to run the same speed-acceleration trace in both MOVES and NeuralMOVES and compare against a chassis dynamometer measurement of a real vehicle, which would reveal whether the surrogate inherits MOVES's biases without its safeguards.","supporting_citations":[{"cited_title":"A., Rodgers, M","cited_arxiv_id":null,"evidence_quote":"MOVES-Matrix, the prior lookup-table approach that NeuralMOVES contrasts with by removing complex inputs and 100 GB storage needs."},{"cited_title":"MOVESTAR: An Open-Source Vehicle Fuel and Emission Model based on USEPA MOVES","cited_arxiv_id":"2008.04986","evidence_quote":"MOVEStar, the prior lightweight MOVES replica that lacks accuracy measures and supports only two vehicle types."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"VT-CPFM, an alternative power-based microscopic model used to show emission profiles and optimal eco-driving trajectories differ from MOVES-based ones."},{"cited_title":"and Wu, C","cited_arxiv_id":null,"evidence_quote":"supplies the reinforcement-learning eco-driving policy used both to generate validation cycles and as the control use case."},{"cited_title":"Congested traffic states in empirical observations and microscopic simulations","cited_arxiv_id":null,"evidence_quote":"IDM, the car-following model that generates intersection-approach validation cycles."}],"review_version":1}