Physics-Informed Predictive Control for Integrated Electric-Vehicle Thermal Management: An Open, Real-Data-Anchored Benchmark

Yifan Wang

arxiv: 2606.22529 · v1 · pith:JTKK4U2Dnew · submitted 2026-06-21 · 📡 eess.SY · cs.SY

Physics-Informed Predictive Control for Integrated Electric-Vehicle Thermal Management: An Open, Real-Data-Anchored Benchmark

Yifan Wang This is my paper

Pith reviewed 2026-06-26 09:45 UTC · model grok-4.3

classification 📡 eess.SY cs.SY

keywords electric vehiclethermal managementmodel predictive controlscientific machine learningbenchmarkbattery thermal managementheat pumpvehicle co-simulation

0 comments

The pith

A shielded physics-informed MPC cuts EV energy use 15 percent while improving comfort and battery health over rule-based control.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper builds an open benchmark, OpenEV-ThermoSciML, that couples a real-data-validated battery electro-thermal-aging model, a two-node cabin model, a heat-pump HVAC model, and a CO2 ventilation model under EPA driving cycles and real weather data. It then trains a physics-informed Sci-ML surrogate that keeps exact conservation laws and uses this surrogate inside a shielded model-predictive controller. Across six scenarios, including a hot-day US06 trip, the controller produces statistically significant gains over a production-style rule-based controller in energy, comfort RMSE, peak cabin CO2, and battery temperature gradient; the same gains appear when the controller is tested on an independently exported OpenModelica eight-node plant.

Core claim

The shielded Sci-ML MPC delivers statistically significant, all-positive improvements over a production-like rule-based controller across six scenarios, including a real hot-day US06 trip (energy −15%, comfort RMSE −47%, peak CO₂ −25%, battery thermal-gradient −78%), and these gains transfer to an independently exported OpenModelica 8-node co-simulation plant.

What carries the argument

Physics-informed Sci-ML surrogate consisting of a nominal-physics prior plus learned residual with explicit conservation penalties, embedded inside a shielded MPC.

If this is right

The open benchmark enables reproducible, multi-objective comparison of thermal-management strategies that previous single-subsystem studies could not provide.
The Sci-ML surrogate remains exact on conserved quantities and shows lower rollout error than black-box or Koopman alternatives out of distribution.
Performance gains hold when the controller is exported to a higher-fidelity eight-node co-simulation plant.
Improvements span battery health, passenger comfort, cabin air quality, and HVAC energy simultaneously.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

If the reduced-order models continue to match higher-fidelity references under varied climates and aging conditions, the same controller structure could be retuned for other battery chemistries or cabin layouts.
The benchmark’s real-data anchoring on BMS logs and cross-check against EnergyPlus suggests it could serve as a shared testbed for comparing future integrated thermal controllers.
Because the surrogate is trained with explicit conservation penalties, similar physics-informed residuals might be added to other vehicle subsystems that exchange heat or mass.

Load-bearing premise

The reduced two-state battery and two-node cabin models remain accurate enough when used inside the MPC that the reported multi-objective improvements will appear in real vehicles.

What would settle it

Run the same hot-day US06 cycle on a physical battery-electric vehicle with the MPC in closed loop versus the rule-based controller and measure whether the four reported percentage improvements in energy, comfort RMSE, peak CO2, and battery thermal gradient are observed.

Figures

Figures reproduced from arXiv: 2606.22529 by Yifan Wang.

**Figure 1.** Figure 1: Independent cabin cross-check. The two-node cabin (orange) already tracks the EnergyPlus single-zone cabin (black) to 1.65 ◦C RMSE; bounded physical re-identification (blue) reaches 0.11 ◦C. Lower panel: pointwise error. 5.2 Closed-loop control: six scenarios (shielded Sci-ML MPC vs rule-based) [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗

**Figure 2.** Figure 2: Closed-loop improvement of the shielded Sci-ML MPC over the rule-based controller on the independent OpenModelica/FMPy plant, by metric and scenario. Every bar is positive. 5.3 Key finding: the cabin–battery co-cooling coupling (scenario E) Scenario E is the physically meaningful coupling point: a 42 ◦C ambient with a hot cabin and a hot pack forces the controller to allocate one heat-pump capacity (the sb… view at source ↗

read the original abstract

Thermal management in a battery-electric vehicle (BEV) is a coupled, vehicle-level problem: the battery pack, the passenger cabin, the heat pump, and cabin air quality compete for shared actuation and energy, yet most studies optimise a single subsystem on proprietary models, which prevents fair, reproducible comparison. We present OpenEV-ThermoSciML, an open and reproducible benchmark that couples a battery electro-thermal-aging model, a two-node cabin model, a heat-pump/HVAC model, and a CO$_2$/ventilation model under real driving cycles (EPA) and real weather (NREL TMY3, NASA POWER), scored by a multi-objective suite spanning battery health, PMV/PPD comfort, cabin air quality, and HVAC energy. The benchmark's battery thermal core is anchored and validated on real BEV battery-management-system (BMS) data; the reduced battery (two-state) and cabin (two-node) models are validated against converged higher-fidelity references and, for the cabin, independently cross-checked against EnergyPlus 25.2.0. On top of the benchmark we develop a physics-informed scientific-machine-learning (Sci-ML) surrogate -- a nominal-physics prior plus a learned residual with conservation penalties -- that is exact on conserved quantities and dominates black-box and Koopman surrogates out-of-distribution (overall rollout RMSE 0.014 vs 1.168 and 3.991). A shielded Sci-ML model-predictive controller (MPC) delivers statistically significant, all-positive improvements over a production-like rule-based controller across six scenarios -- including a real hot-day US06 trip (energy $-15\%$, comfort RMSE $-47\%$, peak CO$_2$ $-25\%$, battery thermal-gradient $-78\%$) -- and these gains transfer to an independently exported OpenModelica 8-node co-simulation plant.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The open benchmark is the real contribution; the MPC transfer claim to the OpenModelica plant lacks the direct closed-loop metric comparison needed to back it up.

read the letter

The standout thing about this paper is the creation of OpenEV-ThermoSciML, an open benchmark that integrates battery electro-thermal-aging, cabin, heat-pump, and CO2 models under real driving and weather data. It is anchored on actual BEV BMS measurements and cross-checked with EnergyPlus, which fills a gap for reproducible work on vehicle-level thermal management.

What the work does well is build a physics-informed surrogate that adds a learned residual to a nominal physics model while enforcing conservation. This surrogate shows much lower rollout error than black-box or Koopman alternatives, especially out of distribution. The shielded MPC then beats a production-like rule-based controller on multiple metrics across six scenarios, including a real hot-day drive, and the authors claim these improvements hold when the controller is tested on an independent OpenModelica plant.

The soft spots are around the closed-loop transfer and model reduction. The abstract states that the gains transfer to the 8-node co-simulation plant, but it does not report direct comparisons of the closed-loop performance numbers between the surrogate-based MPC and the same MPC running on the higher-fidelity plant. Since MPC can amplify small prediction differences into different trajectories, this missing side-by-side data makes the transfer claim harder to assess. The reduced two-state and two-node models are validated in open loop, yet their accuracy when embedded in the optimizer for the multi-objective problem is the key assumption that needs more support.

This paper is for control engineers and thermal modelers who need a shared, data-anchored testbed rather than proprietary single-subsystem studies. A reader who wants to reproduce or extend vehicle thermal MPC work will find the benchmark and validation steps useful. The citation pattern looks reasonable, with references to prior subsystem work and standard tools like EnergyPlus.

It deserves a serious referee. The benchmark itself is a concrete, externally anchored contribution that others can build on, even if the control results would benefit from tighter quantification of plant mismatch.

I would recommend sending it out for peer review.

Referee Report

1 major / 0 minor

Summary. The paper introduces the OpenEV-ThermoSciML open benchmark for coupled BEV thermal management (battery electro-thermal-aging, two-node cabin, heat-pump/HVAC, CO2/ventilation) anchored on real BMS and TMY3 data. Reduced two-state battery and two-node cabin models are validated against higher-fidelity references and EnergyPlus. A physics-informed Sci-ML surrogate (nominal physics plus learned residual with conservation penalties) achieves 0.014 rollout RMSE and outperforms black-box/Koopman alternatives. A shielded Sci-ML MPC is shown to deliver statistically significant all-positive gains over a production-like rule-based controller on six scenarios (e.g., hot-day US06: energy -15%, comfort RMSE -47%, peak CO2 -25%, battery gradient -78%), with the claim that these gains transfer to an independent OpenModelica 8-node co-simulation plant.

Significance. If the transfer claim holds, the work supplies a much-needed open, real-data-anchored, reproducible benchmark that enables fair comparison of multi-objective thermal strategies, addressing the prevalence of proprietary single-subsystem models. The physics-informed surrogate's exact conservation properties and strong out-of-distribution performance, together with the multi-objective MPC results, represent concrete advances for Sci-ML control in coupled vehicle thermal systems.

major comments (1)

[Abstract] Abstract: the assertion that MPC gains transfer to the OpenModelica 8-node co-simulation plant lacks any direct closed-loop metric comparison (energy, comfort RMSE, CO2, battery gradient) between the surrogate-based MPC and the identical MPC deployed on the plant; open-loop rollout RMSE of 0.014 does not establish that prediction discrepancies remain tolerable under the multi-objective optimization and constraints.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the constructive review and for recognizing the benchmark's potential value. We address the single major comment below with a direct response and a commitment to revision.

read point-by-point responses

Referee: [Abstract] Abstract: the assertion that MPC gains transfer to the OpenModelica 8-node co-simulation plant lacks any direct closed-loop metric comparison (energy, comfort RMSE, CO2, battery gradient) between the surrogate-based MPC and the identical MPC deployed on the plant; open-loop rollout RMSE of 0.014 does not establish that prediction discrepancies remain tolerable under the multi-objective optimization and constraints.

Authors: We agree that the abstract's transfer claim would be strengthened by explicit closed-loop metrics on the OpenModelica plant. The current manuscript demonstrates transfer via (i) the surrogate's 0.014 rollout RMSE on held-out trajectories from the higher-fidelity model and (ii) the fact that the MPC formulation is identical when the plant replaces the surrogate. However, we acknowledge that this does not directly quantify how prediction discrepancies propagate through the multi-objective optimizer and constraints. In the revised manuscript we will add a new results subsection (and update the abstract) that reports the four key metrics (energy, comfort RMSE, peak CO2, battery gradient) for the shielded Sci-ML MPC when it is deployed in closed loop on the exported OpenModelica 8-node plant, using the same six scenarios. This will allow direct comparison of surrogate-MPC versus plant-MPC performance. revision: yes

Circularity Check

0 steps flagged

No significant circularity; benchmark anchored externally

full rationale

The paper's central claims rest on external real BMS data for battery anchoring, TMY3/NASA POWER weather, independent EnergyPlus cross-validation for the cabin model, and transfer testing on an exported OpenModelica 8-node plant. The Sci-ML surrogate is validated via open-loop rollout RMSE against these references, and MPC gains are measured against a separate rule-based controller on the benchmark scenarios. No load-bearing step reduces a reported prediction or performance metric to a fitted input by construction, nor does any uniqueness theorem or ansatz trace to self-citation. The derivation chain remains self-contained against the stated external anchors.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review prevents enumeration of specific fitted parameters or axioms; no invented physical entities are mentioned.

pith-pipeline@v0.9.1-grok · 5880 in / 1195 out tokens · 16104 ms · 2026-06-26T09:45:25.214088+00:00 · methodology

discussion (0)

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

Battery thermal-safety reserve erosion by mandatory cabin ventilation in shared-cooling electric vehicles
eess.SY 2026-06 unverdicted novelty 4.0

Mandatory fresh-air ventilation erodes battery thermal reserve in shared-cooling EVs under derated compressor conditions, and a reserve-aware controller using physics-guided ML surrogate and barrier functions restores...

Reference graph

Works this paper leans on

9 extracted references · cited by 1 Pith paper

[1]

drive cycle / vehicle speed→traction powerP trac (road-load equation); 2.P trac +HVAC electrical load→battery current, heat generation, and SOC
[2]

ambient temperature / solar / occupancy→cabin thermal load and CO 2
[3]

HVAC actuation → simultaneouslybattery temperature, cabin temperature, CO 2, and energy
[4]

self-confirming

the controller trades off energy, comfort, CO2/health proxy, battery gradient, and an aging proxy. The single decision that makes the loops genuinely coupled—rather than two independent problems running side by side—is the heat-pumpcapacity split sbatt in §3.3. Allocating capacity to the battery starves the cabin, and vice versa, so coordinated, predictio...
[5]

Anopen, reproducible, weather-coupledwhole-vehicle BEV thermal-management bench- mark unifying battery health, PMV/PPD comfort, cabin air quality, and energy over six standardised scenarios
[6]

Areal-BMS-anchoredbattery thermal core (KU Leuven + NASA POWER; leave-one- vehicle-out validation) and reduced battery/cabin modelsvalidated against converged higher-fidelity references, with the cabin additionally cross-checked against EnergyPlus. 6
[7]

Aphysics-informed Sci-ML surrogatethat is conservation-exact and dominates black-box and Koopman baselines out-of-distribution
[8]

Ashielded, model-swappable multi-objective MPCwith statistically significant, all- positive gains, cross-validated on anindependent OpenModelica FMUclosed loop
[9]

A fully released package (code, processed data, trained models, parameter-provenance and data-mapping documentation) for reproducible comparison. 7 Limitations and future work Validation is in-silico (reduced-order benchmark+ real-data-anchored battery + independent FMU co-simulation + EnergyPlus cabin cross-check);hardware-in-the-loop / bench / vehicle t...

[1] [1]

drive cycle / vehicle speed→traction powerP trac (road-load equation); 2.P trac +HVAC electrical load→battery current, heat generation, and SOC

[2] [2]

ambient temperature / solar / occupancy→cabin thermal load and CO 2

[3] [3]

HVAC actuation → simultaneouslybattery temperature, cabin temperature, CO 2, and energy

[4] [4]

self-confirming

the controller trades off energy, comfort, CO2/health proxy, battery gradient, and an aging proxy. The single decision that makes the loops genuinely coupled—rather than two independent problems running side by side—is the heat-pumpcapacity split sbatt in §3.3. Allocating capacity to the battery starves the cabin, and vice versa, so coordinated, predictio...

[5] [5]

Anopen, reproducible, weather-coupledwhole-vehicle BEV thermal-management bench- mark unifying battery health, PMV/PPD comfort, cabin air quality, and energy over six standardised scenarios

[6] [6]

Areal-BMS-anchoredbattery thermal core (KU Leuven + NASA POWER; leave-one- vehicle-out validation) and reduced battery/cabin modelsvalidated against converged higher-fidelity references, with the cabin additionally cross-checked against EnergyPlus. 6

[7] [7]

Aphysics-informed Sci-ML surrogatethat is conservation-exact and dominates black-box and Koopman baselines out-of-distribution

[8] [8]

Ashielded, model-swappable multi-objective MPCwith statistically significant, all- positive gains, cross-validated on anindependent OpenModelica FMUclosed loop

[9] [9]

A fully released package (code, processed data, trained models, parameter-provenance and data-mapping documentation) for reproducible comparison. 7 Limitations and future work Validation is in-silico (reduced-order benchmark+ real-data-anchored battery + independent FMU co-simulation + EnergyPlus cabin cross-check);hardware-in-the-loop / bench / vehicle t...