{"id":"80af3ec3-12c5-4120-91f8-8bc84c1e0752","arxiv_id":"2411.15929","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A 77-state nonlinear MPC controller was run in real time on an experimental hybrid thermal management testbed, keeping the cold plate below 45°C while using phase-change storage to shave transient heat loads.","lead":"This paper demonstrates a nonlinear model predictive controller with 77 states running in real time on an experimental vehicle thermal management system with phase-change thermal energy storage. The method linearizes the stiff thermal dynamics at each control step, letting one controller update finish in about half a second.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The gradients supplied to fmincon (Eq. 20) do not follow from Eqs. (18)-(19), and with r=0 the inverse is only warm-started; without a check against exact gradients or exact integration, the claimed 1 s NMPC solution is not established.","rationale":"The paper's strongest claim is the first experimental real-time implementation of a 77-state NMPC on a hybrid thermal management system. I read the construction in good faith: the graph-based model, the linearized trapezoidal integration step, and the warm-started Newton-Schulz inverse are plausible mechanisms for making a large stiff prediction model cheap to evaluate, and the reported execution times and experimental temperature traces support the claim that some predictive controller ran in real time. What must be true for the label 'NMPC' is that fmincon is actually optimizing the discretized optimal control problem given by Eq. (7) with the integrator of Algorithm 1. That requires the analytic gradient to be the derivative of that objective with respect to the decision variables. The derivation in Section IV-B does not deliver this: Eq. (19) and Eq. (20) are mutually inconsistent, and the derivative of the approximate inverse with respect to u is dropped even though r=0 makes that inverse a stale warm start rather than an accurate function of the current decision variables. This is an internal correctness issue rather than a disagreement with community consensus. I did not select the single-run no-TES comparison as the load-bearing concern: it weakens the quantitative 7 C claim and the absence of error bars is real, but the first-implementation claim and the numerical contribution are more central. The reader's weakest assumption identified the same approximate-gradient and one-iteration inverse issue, so my verdict is unchanged: conditional, because the numerical claims need direct verification before the method can be credited as a real-time NMPC solution.","tokens_in":13371,"tokens_out":9117,"duration_ms":82960,"concrete_test":"Recompute the rollout defined by Algorithm 1 with r=0 for the Fig. 3 scenario and evaluate dJ/dU by central finite differences or automatic differentiation of the actual objective implemented in the paper; compare this exact Jacobian to Eq. (20) at several time steps k. If the relative error exceeds 10% or the search direction is wrong, rerun fmincon with the exact Jacobian and compare the cold-plate trajectory, constraint violations, and per-step iteration count. If the exact-Jacobian run produces materially different control inputs or exceeds the 1 s update budget, the reported controller is not the claimed NMPC solution and the conditions of the paper need to be revised.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that a 77-state NMPC is solved online at 1 s rests on the analytic gradients given to fmincon and on the accuracy of the r=0 Newton-Schulz inverse. Neither is validated in the paper. In Section IV-B, Eq. (19) states that the derivative quantity on the left is approximately t_s exp(t_s A_k) G_j,k, and Eq. (18) would then give partial x_{k+1}/partial u_j,k approximately t_s exp(t_s A_k) G_j,k x_k. Instead Eq. (20) gives t_c A_k E_k [I_nu tensor x_k]. These differ already in the leading term, which should be t_s G_j,k x_k, not t_c A_k G_j,k x_k. Moreover, Eq. (20) is not the exact derivative of the rollout in Eq. (14), because the dependence of the approximate inverse D_k^{-1} on u_k is ignored. This omission matters precisely when r=0, since D_k^{-1} is then just the previous step's inverse. No finite-difference gradient comparison, SQP iteration count, or KKT residual is reported, so it is unknown whether fmincon actually reaches a stationary point of the discretized optimal control problem. The experimental traces may still demonstrate a fast heuristic controller that uses TES and keeps the cold plate cooler, but the transferable numerical method and the claim of solving NMPC online are not yet supported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents a nonlinear model predictive controller (NMPC) for a hybrid thermal management system (TMS) with latent thermal energy storage (TES), based on a 77-state model (5 lumped fluid/component states plus 72 TES states). The authors propose an explicit integration method that linearizes the dynamics at each time step within the MPC horizon and applies a trapezoidal rule with a Newton-Schulz matrix inverse approximation. They further supply analytic gradients to the MATLAB solver fmincon. The paper reports simulation and experimental results on a laboratory testbed showing that the controller meets the cold-plate temperature objective, and a comparative experiment in which the hybrid TMS maintains a cold-plate temperature about 7 °C lower than a TMS without TES under the same transient heat load profile.","tokens_in":13672,"tokens_out":7380,"duration_ms":67611,"significance":"If the numerical methodology were fully validated, this would be a notable demonstration: real-time NMPC with a high-dimensional, stiff prediction model on an experimental hybrid TMS, and the first experimental MPC implementation on such a system. The physical experiment comparing with and without TES provides concrete evidence of the benefit of actively controlling TES charging and discharging, which is a valuable contribution. However, the central numerical claims are not yet supported as stated. The analytic gradient formula in Eq. (20) appears algebraically inconsistent with Eqs. (18)-(19), the r=0 Newton-Schulz approximation is used without error analysis, and there are no convergence diagnostics or comparisons against exact gradients or a stiff integrator. These gaps directly affect the claim that the paper solves the NMPC optimal control problem online rather than implementing a heuristic controller. The experimental results are encouraging but are based on a single trial per configuration, so the quantitative benefit claim should be treated cautiously.","major_comments":[{"comment":"The gradient formula in Eq. (20) does not follow from Eqs. (18) and (19). From Eq. (19), the derivative of the state transition with respect to u_{j,k} is approximately t_s exp(t_s A_k) G_{j,k}, whose leading term is t_s G_{j,k}; combined with Eq. (18) this gives a leading term t_s G_{j,k} x_k. Eq. (20) instead gives t_c A_k G_{j,k} x_k, which differs already in the leading order. Furthermore, Eq. (20) ignores the dependence of D_k^{-1} in Eq. (14) on u_k; this dependence is present whenever the Newton-Schulz iteration runs for r>0, and even for r=0 the first horizon step uses an initial inverse computed from D_0, which depends on u_0. Since the optimizer is configured with 'Specify Objective Gradient True', incorrect gradients mean fmincon may not converge to a stationary point of the discretized optimal control problem. The authors should correct Eq. (20) or, if they intend a different approximation, explain it precisely and validate it against finite-difference gradients or the exact derivative of the rollout, reporting convergence measures such as SQP iteration counts, KKT residuals, or first-order optimality values.","section":"Section IV-B, Eq. (20)"},{"comment":"The Newton-Schulz matrix inverse approximation is used with r=0 (one iteration) and the previous step's inverse as the initial guess. The paper provides no error analysis for this approximation: no matrix inverse residual, condition number, or comparison of the resulting rollout against a standard stiff solver such as ode15s is reported. Given that the same approximate inverse is used across the horizon and also feeds the gradient computation, the accuracy of both the prediction and the optimization is unverified. Please report integration error statistics over the prediction horizon and demonstrate that one Newton-Schulz iteration is sufficient under the tested operating conditions.","section":"Section IV-A, Algorithm 1, and Table III"},{"comment":"The quantitative claim that the hybrid TMS maintains the cold plate approximately 7 °C lower than the TMS without TES is based on a single experimental run for each configuration. There are no repeated trials, no error bars, and no discussion of thermocouple measurement uncertainty or run-to-run variability in the initial condition. As a proof-of-concept the single trial is informative, but the quantitative comparison should be framed as a single-trial observation and the measurement uncertainty should be stated.","section":"Section V-B3, Fig. 11"},{"comment":"Eq. (17) states that ∂x_{k+1}/∂x_k is approximately A_k, but the actual state transition of the linearized trapezoidal scheme in Eq. (14) is D_k^{-1}(I + (t_c/2)A_k) (with the appropriate augmented block). The approximation A_k is not justified and propagates through the chain rule in Eq. (16), affecting the total gradient. The authors should use the exact transition matrix of the linearized scheme or provide evidence that A_k is an adequate approximation over the horizon.","section":"Equation (17)"}],"minor_comments":[{"comment":"The text 'mimicking a head load generated by high power electronics' should read 'heat load'.","section":"Section II-A"},{"comment":"The phrase 'F or i∈...' contains a typo, and the symbol A_k is reused both for the dynamics matrix in Eq. (6) and for the top-left block of the state transition matrix in Algorithm 1; these should be distinguished.","section":"Algorithm 1"},{"comment":"The statement 'twelve of the seventy-two TES states are directly measured' appears inconsistent with the subsequent description that for each module the fluid inlet, fluid outlet, and three PCM states are used with the observer, which would imply twenty measured quantities per four modules. Please clarify which TES states are directly measured.","section":"Section V-B1"},{"comment":"The sentence 'the value of t he function is zero' contains a typo and should read 'the value of the function'.","section":"Section V-B2"},{"comment":"The execution-time plot would be clearer if it also reported the maximum, mean, and standard deviation of computation time over the run, rather than relying solely on visual inspection of the trace.","section":"Fig. 7"}],"recommendation":"major_revision","confidential_remarks":"The paper's claim to be the first experimental MPC implementation for a hybrid TMS is plausible but not independently verified here. The main technical risk is the unvalidated gradient and integration approximations; if the authors can correct Eq. (20) and provide numerical validation (gradient checks, integration error, convergence diagnostics), the contribution would be strong. The experimental comparison with and without TES is valuable even if the numerical method is presented as heuristic, but the current framing overstates the optimization guarantee. I recommend major revision rather than rejection because the experimental demonstration itself appears sound and the numerical issues are addressable within the manuscript's scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is the first experimental real-time NMPC on a hybrid TMS with 77 states and a 1 s update, and the comparison against a no-TES version is a real physical experiment showing ~7°C lower cold plate temperature. That part is worth something. The paper also gives a clear description of the testbed, the observer setup, and the cost design. Credit where due: the execution times in Fig. 7 are below the sample period, the temperature traces match simulation reasonably well, and the TES charging/discharging behavior is sensible.\n\nThe soft spot is Section IV-B, and it is not minor. The approximate gradient formula, Eq. (20), does not follow from Eq. (18)-(19). Eq. (18) says the state-to-input derivative is (∂A_k/∂u_j,k) x_k, which is itself an approximation that ignores the dependence of the inverse D_k^{-1} on u_k. Eq. (19) then writes ∂A_k/∂u_j,k ≈ t_s exp(t_s A_k) G_j,k, which is a matrix exponential times G, not a derivative of A. And Eq. (20) gives t_c A_k E_k [I⊗x_k], which doesn't match the leading term t_s G_j,k x_k that you would get from the matrix exponential derivative. On top of that, with r=0 the Newton-Schulz inverse is simply the previous step's inverse, so the integration itself is also approximate in a way that is not analyzed. No comparison against exact gradients, finite differences, or exact integration is reported, so it is unknown whether fmincon is actually minimizing the stated OCP or just a nearby heuristic. The experimental traces may still show a working controller—it kept the cold plate cool and the TES helped—but the transferable numerical method and the 'online NMPC' claim are not yet established as written.\n\nI would not desk-reject this. The experimental demonstration is novel and the TES benefit is physically meaningful. A serious referee could ask for a corrected gradient derivation, a finite-difference check on the gradients, and at least one repeat of the experiment or a sensitivity analysis. If the authors can show the approximate gradients are close to exact, the paper becomes solid. As is, it's a good engineering demo with a broken appendix in the middle.\n\nRecommendation: send to peer review, but expect heavy revision on Section IV.","headline":"A real 77-state NMPC on a hybrid thermal management testbed, with a genuinely useful TES-vs-no-TES experiment—but the gradient derivation in Section IV-B does not hold up, so the 'solved online' claim is only partially supported.","tokens_in":14206,"tokens_out":3517,"would_cite":false,"duration_ms":29120,"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":"Nonlinear model predictive control can run in real time on a stiff, 77-state hybrid thermal management system: the paper demonstrates the first experimental MPC implementation on such a testbed, with a 25-step horizon updating every…","keywords":["nonlinear model predictive control","thermal management system","latent thermal energy storage","phase change material","stiff dynamics","real-time implementation","matrix inverse approximation","state estimation"],"falsifier":"Rerun the same simulated and experimental scenarios with the matrix inverse computed exactly, or with two or more Newton-Schulz iterations, and with exact gradients; if the control inputs, cold-plate trajectories, or execution times change substantially, the approximations are doing nontrivial work and the claim that they are sufficiently accurate is in doubt. A second test is to apply a heat-load profile with a fast, large pulse that pushes the cold plate near its 45 °C limit and check whether the 1-second update rate still holds.","tokens_in":13129,"feed_emoji":"🧊","tokens_out":5328,"duration_ms":45751,"temperature":0.7,"pith_summary":"The paper sets out to show that nonlinear model predictive control can run in real time on a high-dimensional, stiff thermal management system with integrated latent thermal energy storage. It reports the first experimental implementation of MPC, linear or nonlinear, on such a hybrid system: a 77-state controller updated every second on a physical testbed. The authors argue that linearizing the state-dependent dynamics at each prediction step turns stiff integration into an explicit, low-cost computation and yields analytic gradients almost for free. If true, this removes a practical barrier to using MPC where fine-grained phase-change models are needed, and it gives a concrete way to test the value of actively charging and discharging phase-change storage under transient heat loads.","feed_headline":"First real-time nonlinear MPC on a hybrid thermal testbed","feed_subtitle":"A 77-state controller updates every second and keeps the cold plate about 7 °C cooler with thermal storage.","key_machinery":"The mechanism that carries the argument is the graph-based state-dependent coefficient model, $M(x)\\dot{x}=C(x,u,d)x+Bd$, where $M$ is a diagonal thermal-capacitance matrix and $C$ is a sparse thermal-conductance matrix. Linearizing at each MPC time step turns the prediction model into $\\dot{x}=A_k x+\\tilde{d}_k$, whose trapezoidal-rule integration is $(I-\\frac{t_c}{2}\\phi_k)^{-1}(I+\\frac{t_c}{2}\\phi_k)\\tilde{x}_k$; the inverse is computed by a fixed number of Newton-Schulz iterations, starting from the previous step's inverse, which makes the whole integrator a fixed sequence of matrix operations. The same linearized form yields a cheap first-order gradient approximation, $\\partial x_{k+1}/\\partial u_k \\approx t_c A_k E_k [I_{n_u}\\otimes x_k]$, so the optimizer gets analytic gradients without finite differences. This combination is what fits 77 states and 25 prediction steps inside a 1-second update interval.","core_discovery":"The central claim is that a hybrid thermal management system—a pumped coolant loop with a cold plate, a heat exchanger, and four finned phase-change storage devices—can be controlled by a nonlinear MPC that solves a 25-step optimal control problem in under one second, despite having 77 states and dynamics with widely separated timescales. The enabler is a state-dependent graph model of the form $\\dot{x}=A(x,u,d)x+\\tilde{d}$; fixing $A$ at each step gives a linear system whose implicit trapezoidal integration has an explicit closed form, with the required inverse approximated by Newton-Schulz iteration. The same linearization provides one-term power-series approximations to the gradients, so a standard sequential quadratic programming optimizer receives analytical gradients at minimal extra cost. The authors validate this on an experimental testbed, matching simulated trajectories, and report that adding the thermal storage keeps the cold plate roughly 7 °C cooler than an otherwise comparable MPC without storage under the same transient heat loads. They also state that this is the first experimental demonstration of MPC, linear or nonlinear, for a hybrid thermal management system.","pith_inferences":["Editorial inference: the 7 °C benefit is tied to this specific load profile and to the sizing of the storage and heat exchanger; nothing in the paper claims it generalizes to all loads, and co-designing the storage volume could make it larger or smaller.","Editorial inference: the approximations form a tunable family—more Newton-Schulz iterations, higher-order gradient terms, or exact inverse computation—so the accuracy-versus-update-rate frontier could be mapped explicitly, which the paper does not do.","Editorial inference: the graph-based state-dependent representation is modular, so the same machinery could extend to other phase-change or transport-reaction systems, but faster dynamics than the cold plate would stress the one-point linearization assumption.","Editorial inference: separating the contribution of the integrator from the contribution of the gradient shortcut would clarify which piece is responsible for the computational savings, since the paper validates them together."],"forward_implications":["Other high-dimensional, stiff thermo-fluid systems with state-dependent coefficients can use the same explicit-integrator-plus-analytic-gradient recipe to run nonlinear MPC at update rates that were previously out of reach.","Active charging and discharging control of phase-change storage becomes testable on hardware rather than only in simulation, giving system designers a direct measurement of how much thermal endurance a given storage volume provides.","Because the cost function penalizes the TES temperature states, the controller naturally allocates flow to recharge the storage during low-load periods, so thermal endurance is actively scheduled rather than left to passive behavior.","The experimental comparison with a no-TES MPC quantifies the benefit of hybridization on this testbed: about 7 °C lower cold-plate temperature during the tested transient profile, at roughly similar primary flow rates.","The fixed iteration count of the integrator makes worst-case computation time predictable, which is a useful property for safety-critical embedded control applications."],"supporting_citations":[{"why":"Supplies the finite-volume TES model and the nonlinear observer used to estimate the 60 unmeasured TES states during the experiment.","marker":"[9]"},{"why":"Provides the stability-function treatment of implicit Runge-Kutta methods that justifies using the trapezoidal rule for stiff integration.","marker":"[11]"},{"why":"Supplies the governing lumped-parameter equations for the tank, cold plate, and heat exchanger restated in Eq. (1).","marker":"[15]"},{"why":"Supports the composite PCM-metal control-volume model with effective material properties used to discretize the TES devices.","marker":"[16]"},{"why":"Supports the effective-specific-heat function that lets melt fraction be computed from control-volume temperatures.","marker":"[17]"},{"why":"Introduces the iterative Newton-Schulz method for approximating the matrix inverse used inside the integrator.","marker":"[20]"},{"why":"Provides the improved Newton iteration for the matrix inverse that underlies the approximate inversion step in Algorithm 1.","marker":"[21]"},{"why":"Supplies exact derivatives of the matrix exponential that the one-term power-series gradient approximation draws on.","marker":"[24]"},{"why":"Is the source of the first-term power-series approximation to the gradient used in Eq. (20).","marker":"[25]"}],"fun_headline_variants":["77-state nonlinear MPC solves in under a second on thermal rig","Real-time nonlinear MPC: 77 states, sub-second control","Sub-second nonlinear MPC for 77-state hybrid thermal system","Nonlinear MPC handles 77 stiff states in under a second","First experimental nonlinear MPC for hybrid thermal management"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole real-time scheme rests on the assumption that one Newton-Schulz iteration for the matrix inverse and a one-term power-series formula for the gradients are accurate enough for the optimizer: the controller's feasibility and the reported 7 °C benefit could fail to generalize if those approximations silently bias the solution.","fun_headline_variants_meta":{"raw":{"variants":["77-state nonlinear MPC solves in under a second on thermal rig","Real-time nonlinear MPC: 77 states, sub-second control","Sub-second nonlinear MPC for 77-state hybrid thermal system","Nonlinear MPC handles 77 stiff states in under a second","First experimental nonlinear MPC for hybrid thermal management"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000833,"raw_usage":{"total_tokens":3661,"prompt_tokens":993,"completion_tokens":2668,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":609,"completion_tokens_details":{"reasoning_tokens":2587}},"tokens_in":609,"tokens_out":2668,"duration_ms":19054,"temperature":1.0,"reasoning_tokens":2587,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T13:44:22.860411+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Rerun the same simulated and experimental scenarios with the matrix inverse computed exactly, or with two or more Newton-Schulz iterations, and with exact gradients; if the control inputs, cold-plate trajectories, or execution times change substantially, the approximations are doing nontrivial work and the claim that they are sufficiently accurate is in doubt. A second test is to apply a heat-load profile with a fast, large pulse that pushes the cold plate near its 45 °C limit and check whether the 1-second update rate still holds.","supporting_citations":[{"cited_title":"Design and Vali- dation of a State-Dependent Riccati Equation Filter for State of Charge Estimation in a Latent Thermal Storage Device,","cited_arxiv_id":null,"evidence_quote":"Supplies the finite-volume TES model and the nonlinear observer used to estimate the 60 unmeasured TES states during the experiment."},{"cited_title":"Hairer and G","cited_arxiv_id":null,"evidence_quote":"Provides the stability-function treatment of implicit Runge-Kutta methods that justifies using the trapezoidal rule for stiff integration."},{"cited_title":"Control of a Hybrid Thermal Man- agement System: A Heuristic Strategy for Charging and Dis- charging a Latent Thermal Energy Storage Device,","cited_arxiv_id":null,"evidence_quote":"Supplies the governing lumped-parameter equations for the tank, cold plate, and heat exchanger restated in Eq. (1)."},{"cited_title":"Conductive heat transfer in lamellar phase change material composites,","cited_arxiv_id":null,"evidence_quote":"Supports the composite PCM-metal control-volume model with effective material properties used to discretize the TES devices."},{"cited_title":"Numerical Validation of Effective Specific Heat Functions for Simulating Melting Dynamics in Latent Heat Thermal Energy Storage Modules,","cited_arxiv_id":null,"evidence_quote":"Supports the effective-specific-heat function that lets melt fraction be computed from control-volume temperatures."},{"cited_title":"Iterative Berechung der reziproken Matrix,","cited_arxiv_id":null,"evidence_quote":"Introduces the iterative Newton-Schulz method for approximating the matrix inverse used inside the integrator."},{"cited_title":"An Improved Newton Iteration for the Generalized Inverse of a Matrix, with Applications,","cited_arxiv_id":null,"evidence_quote":"Provides the improved Newton iteration for the matrix inverse that underlies the approximate inversion step in Algorithm 1."},{"cited_title":"Derivatives of the Matrix Expo- nential and Their Computation,","cited_arxiv_id":null,"evidence_quote":"Supplies exact derivatives of the matrix exponential that the one-term power-series gradient approximation draws on."},{"cited_title":"Random-effects substitution models for phylogenetics via scalable gradient approximations,","cited_arxiv_id":null,"evidence_quote":"Is the source of the first-term power-series approximation to the gradient used in Eq. (20)."}],"review_version":1}