Physics-Informed CNN for Battery Pack Temperature Fields

New
License: arXiv preprint (copyright held by authors) · Updated 2026-11-16
My rating

1. Overview

Battery thermal management is a good test case for physics-informed surrogates because the dominant cost is not the network but the labelled data. Generating a temperature field for each design variant requires a finite-element or CFD solve, and a purely data-driven model needs many of them. The preprint arXiv:2411.09915 addresses exactly that bottleneck: it builds a 21700 cell pack with indirect liquid cooling, calibrates a simplified finite-element model against experiment, and trains a physics-informed convolutional neural network whose loss includes a finite-difference discretisation of the heat conduction equation. The reported result is an accuracy improvement of more than fifteen percent over a data-driven model trained on the same data.

This entry is a case study, not a product review. It explains what the physics term in the loss actually enforces, which assumptions make the surrogate tractable, how the data chain is built from experiment to finite-element corpus to network, and what the reported improvement does and does not establish. The audience includes engineers evaluating a physics-informed approach for a thermal design loop, reviewers checking a claim that a surrogate will cut simulation cost, and readers of the PINN landscape guide who want a concrete industrial example. The classical solver cost guide supplies the amortisation arithmetic that this case deliberately leaves open.

Scope limits matter here. The source is a preprint, not a released codebase or a certified validation report. The abstract does not specify the accuracy metric behind the fifteen-percent figure, nor an absolute error in kelvin, nor hardware and training budgets. This guide therefore treats the number as a directional result from the authors' own comparison and does not convert it into a speed-up or a cost saving. It also does not claim the approach generalises to other cell formats, cooling topologies or drive cycles without new experiments. All source facts were checked on 9 November 2026 against the arXiv record and are dated in Section 9.

2. How It Works

The physical system is a battery pack built from 21700 cylindrical cells with indirect liquid cooling. Cold plates sit above and below the cells, and thermal paste surrounds the cells to provide a conduction path. Because the coolant flow rate is high, the authors treat the cold plates as constant-temperature boundaries and model the cells as heat sources. That simplification is what makes a finite-element training corpus affordable: the fluid side of the conjugate heat transfer problem is replaced by a fixed boundary condition, and the remaining problem is heat conduction in a solid with distributed sources. A simplified finite-element model is then calibrated against experimental measurements before it is used to generate data.

The surrogate is a convolutional neural network that outputs a temperature field. A conventional data-driven loss penalises the discrepancy between predicted and labelled fields. The physics-informed loss adds a residual constructed from the heat conduction equation discretised by finite differences, so a prediction is penalised for violating local energy balance even where no labelled value exists. In operator form the composite objective has a data term and a physics term with a relative weight, just as in the broader PINN family described in the PINN landscape guide. The finite-difference formulation is notable: derivatives are computed on the output grid with a fixed stencil rather than by automatic differentiation of a continuous network, which suits a convolutional field predictor and keeps the residual cheap to evaluate.

The claimed benefit follows from that extra supervision. With fewer labelled simulations, the residual term constrains the space of admissible fields and helps training converge. The authors report that the physics-informed network improves accuracy by more than fifteen percent relative to a data-driven method trained on exactly the same dataset. The comparison is controlled on the data axis, which is the right control for a claim about label efficiency. It is not controlled on the compute axis: evaluating the residual adds work per training step, and the abstract does not report it. A fair project-level comparison would count both the saved finite-element solves and the added training cost, following the four-ledger method in the cost boundary guide.

3. Scope and Assumptions

The constant-temperature cold-plate assumption is the most consequential simplification. It is justified in the preprint by a high coolant flow rate, and it removes the need to resolve coolant flow, temperature rise along the channel, and conjugate heat transfer at the plate. If a design changes the flow rate, channel geometry or coolant inlet temperature, the assumption must be re-examined; a surrogate trained under it has learned a boundary condition, not the full thermal system. Cell heat generation is treated as a source term, and the finite-element model is simplified and calibrated to experiment, so its validity is tied to the operating range of that calibration.

The discretisation assumption is equally concrete. Writing the heat conduction residual with finite differences ties the physics term to the grid spacing and stencil used for the output field. Change the resolution and the residual changes meaning; change the units and the relative weight between data and physics terms changes. Anyone reproducing this case must record grid spacing, time discretisation if transient, material properties including anisotropy, contact resistances at the thermal paste interfaces, and the source-term model. These are the quantities that turn a research result into a reusable thermal case rather than a one-off figure.

The data regime is supervised simulation data with a physics regulariser. It is not a sensor-only setting, and it is not a zero-label PDE solve: labelled finite-element fields are still used. The PINN landscape guide explains why pure residual training can collapse to trivial solutions on some problems and why sparse measurements are a standard design element. The licence and availability assumption also needs stating: arXiv:2411.09915 is a preprint with copyright held by the authors. This entry does not claim a public code release, a validated dataset or a permissive software licence. Any commercial reuse requires checking the actual artefact and its terms, not the catalogue card.

4. Getting Started

A practical reproduction starts from the physical setup, not from the network. Define the pack geometry, the number and arrangement of 21700 cells, the cold-plate positions and the thermal paste layers. Establish the boundary condition that represents the plates and the source term that represents cell heat generation. Build the simplified finite-element model, calibrate it against measured temperatures, and only then generate a corpus of labelled temperature fields. Tools such as FEniCS or a general CFD solver such as OpenFOAM can supply the forward solver; PyBaMM is the natural choice when the electrochemical side must be modelled explicitly. This order matters because calibrating on a handful of experiments and then sweeping parameters is the whole cost argument.

The network side can be implemented in any framework that supports a convolutional field predictor and a differentiable finite-difference stencil. DeepXDE and PhysicsNeMo provide physics-informed training infrastructure, PINNs from scratch covers the residual mechanics, and the PINN getting-started collection gathers the surrounding learning material. The composite loss needs a weighting strategy: too small a physics weight gives no label-efficiency benefit, too large a weight can dominate the data fit and bias the field toward an oversimplified conduction model. A small ablation over the weight, with a fixed data budget, is a sensible first experiment.

Before trusting the result, run three sanity checks. First, train a purely data-driven baseline on the identical corpus and compare with the same metric, which is the control the preprint uses. Second, sweep the number of labelled simulations downward and report accuracy as a function of corpus size; a label-efficiency claim is a curve, not a single point. Third, evaluate on operating conditions outside the finite-element calibration range, including a higher heat-generation rate and a different boundary temperature. Record absolute error in kelvin, peak temperature error and the location of the hot spot, because a low field-average error can hide a safety-critical local miss. The surrogate modelling collection and PDEBench give useful templates for reporting such curves.

5. Comparison with Alternatives

The natural comparator is the purely data-driven CNN trained on the same labelled fields. That is the baseline in the preprint and the cleanest way to isolate the value of the physics term. A second comparator is a classical reduced-order model: proper orthogonal decomposition with an interpolated coefficient vector is cheap, auditable and often adequate when the temperature field varies smoothly with a few design parameters. A third is a full finite-element solve with no surrogate, which remains the reference for accuracy and is the correct choice for a one-off design check. The cost boundary guide frames the choice as a break-even query count rather than an architecture preference.

Within physics-informed methods, the contrasts are informative. A 2025 study on lithium-ion temperature prediction couples an enhanced Bernardi heat-generation model with a PINN for a 50 Ah prismatic cell, uses twelve surface thermocouples and reports sub-kelvin errors under constant-current and drive-cycle conditions; that is a sensor-driven field-reconstruction problem, with different data and a different cell format. A 2026 line of work on fully coupled electrochemical-thermal pseudo-2D models uses log-scale parameterisation and automatic loss balancing to keep the residual terms stable. A five-tier digital twin framework combines PyBaMM simulations, PhysicsNeMo PINNs and uncertainty quantification. These are related techniques, not alternative numbers for the same benchmark, and their reported errors should not be placed in one table as if they were comparable.

The physics-informed CNN approach also differs from neural operator learning. An operator model learns a map from coefficients or forcing to a solution field across a family of problems and needs a large paired corpus; this case trains a field predictor for one pack configuration with a residual regulariser. If the deliverable is a design sweep over many cooling layouts, an operator or a reduced-order model may be more appropriate. If the deliverable is label efficiency within one layout, the physics-informed loss is the relevant lever. Stating which question is being asked prevents an apples-to-oranges comparison.

6. Benchmarks and Evidence

The only quantitative result in the preprint abstract is that the physics-informed CNN improves accuracy by more than fifteen percent over a data-driven method trained on the same training data. That is a relative, author-reported comparison. The abstract does not name the accuracy metric, does not give an absolute error in kelvin, does not report the corpus size, does not state the hardware or training budget, and does not say whether the improvement is an average over conditions or a specific operating point. It is therefore a directional result about label efficiency, not a validated accuracy specification for a production thermal model.

A stronger evidence package would include the following. Report the number of finite-element simulations used for training and testing, and plot error against corpus size for both the physics-informed model and the data-driven baseline. Report absolute error and peak-temperature error in kelvin, with the hot-spot location, because thermal design is usually constrained by a maximum rather than a mean. State the physics-loss weight and its sensitivity. Report residual-evaluation overhead per training step so that the label saving can be converted into a wall-clock or core-hour saving. Finally, report an out-of-calibration test with a different heat-generation rate or boundary temperature, which is where a physics regulariser either helps or quietly fails.

Contextual numbers from adjacent work must stay in their own column. The prismatic-cell PINN study reports sub-kelvin errors for a 50 Ah ternary cell under constant-current and UDDS conditions with twelve thermocouples; the digital twin framework reports calibration errors for voltage and temperature against a battery dataset; neither uses the same pack, corpus, boundary condition or metric as arXiv:2411.09915. The classical solver cost guide gives the template for separating these clocks and quantities of interest. None of the sources above supplies a certified validation dataset for thermal runaway or safety envelopes, so no claim about certification is made here. Any proposal number should come from a local reproduction with the case setup, corpus, loss weight and hardware archived alongside the measurement date.

7. Role in Physics-AI Workflows

In a physics-AI pipeline this case occupies the "surrogate inside a design loop" slot. The finite-element model is the forward solver and corpus generator; the physics-informed CNN is the fast evaluator; the engineering consumer is a cooling-design sweep or a parameter study. The value proposition is that fewer finite-element solves are needed to reach a usable accuracy, which shifts the fixed cost term in the amortisation formula. The cost boundary guide writes that term as the sum of corpus generation, training and maintenance, divided by the per-query saving. A label-efficiency result lowers one term; it does not by itself make the whole project pay off.

The case also illustrates how physics constraints interact with data provenance. The residual term is derived from the same simplified conduction model that produced the labels, so it cannot correct a systematic bias in that model. If the constant-temperature boundary is wrong for a new operating point, both the labels and the physics term agree on the wrong answer. A useful deployment therefore keeps a verification path to a higher-fidelity solve or to experiment, with an explicit trigger for when the surrogate is not trusted. The PINN landscape guide discusses the same issue under residual-based training and sparse observation guidance.

For adjacent battery workloads, the site's PyBaMM battery twin case covers electrochemical-thermal coupling, and the chip thermal case shows a physics-informed surrogate in a different heat-transfer setting. Building energy surrogates are another example of a reduced-order model with a well-defined quantity of interest. Together these entries show the pattern: the surrogate is valuable when queries repeat inside a validated envelope, and the solver or experiment stays in the loop as the reference. The physics-AI learning path places this pattern in a broader sequence.

8. Limitations and Failure Modes

The boundary simplification is the first failure point. Constant-temperature cold plates follow from a high coolant flow rate in the reported setup; at lower flow rates, with a different channel layout, or when the coolant temperature rises appreciably along the plate, the assumption breaks. The surrogate has learned a field consistent with that boundary, and no physics term derived from the same model will reveal the discrepancy. A change in cell format, contact resistance at the thermal paste, or anisotropic conductivity has a similar effect: the labels and the residual move together, so validation must be external.

Residual-based training has its own pathologies. If the physics weight dominates, the network can fit a smooth conduction field that satisfies the discretised equation while missing the measured hot spot. If it is too small, the method degenerates to the data-driven baseline. Finite-difference residuals are grid-dependent and sensitive to units and scaling, so a change in output resolution silently changes the objective. The PINN landscape guide catalogues gradient conflict, magnitude imbalance and trivial-solution collapse in the broader family; a convolutional field predictor with a fixed stencil avoids some of them but not all. Long-horizon or transient extrapolation is also unverified: the abstract describes a temperature field surrogate, not a validated transient thermal model over a drive cycle.

Finally, there are delivery limits. The source is a preprint with no stated public code or data release, so reproducibility depends on re-implementing the setup. The reported improvement has no absolute error, no uncertainty estimate and no hardware budget. Thermal design often feeds safety analysis, where peak temperature and its location matter more than a field average and where certification requires documented verification. When the deliverable is a safety envelope, a fresh high-fidelity solve or an experiment remains the reference and the surrogate can only be a screening tool. The cost boundary guide explains why a deferral rule back to the solver belongs in the design, and the neural operator selection guide lists the out-of-distribution stress tests that any learned field model should pass before deployment.

9. Version and Source Notes

Record 1: arXiv:2411.09915 is titled "Physics-informed Machine Learning for Battery Pack Thermal Management" by Zheng Liu, Yuan Jiang, Yumeng Li and Pingfeng Wang, submitted to cs.LG with v1 on 15 November 2024 and v2 on 18 November 2024. Both dates were checked against the arXiv record on 9 November 2026. The catalogue year value of 2024 refers to that preprint, not to a software release or a journal publication. Record 2: the abstract describes a 21700 pack with indirect liquid cooling, cold plates above and below, thermal paste around the cells, a simplified finite-element model calibrated against experiment, and cold plates treated as constant-temperature boundaries because of the high coolant flow rate.

Record 3 concerns an earlier catalogue description on this site, which claimed a PyTorch/TensorFlow multi-backend implementation suitable as a first framework for physics AI. The preprint abstract does not support a multi-backend code release or a general framework claim; it reports a research surrogate for one pack configuration. This upgraded entry replaces that description with the narrower, source-backed scope. The correction is recorded here rather than silently dropped, because a catalogue card and a case study are different evidence classes. Record 4: the "more than fifteen percent" figure is an author-reported relative accuracy improvement over a data-driven baseline trained on the same data; the abstract does not specify the metric, an absolute error or a compute budget, so this guide treats it as directional.

Record 5: the finite-element model is simplified and calibrated, and the physics loss is built from a finite-difference discretisation of heat conduction. Both are model choices, not first-principles guarantees. Record 6: adjacent 2025 and 2026 studies on prismatic cells, P2D coupling and digital twins use different cell formats, boundary conditions, sensors and metrics; their numbers are contextual and are not merged into a single comparison. Record 7: arXiv preprints carry copyright held by the authors, so no permissive code licence is asserted. Where these records conflict with a broader public claim, this guide keeps the narrower version and attributes the wider claim to its source.

10. References

  1. Liu Z., Jiang Y., Li Y., Wang P., "Physics-informed Machine Learning for Battery Pack Thermal Management", arXiv:2411.09915, v1 15 Nov 2024, v2 18 Nov 2024 (accessed 9 November 2026).
  2. arXiv record and abstract for 2411.09915, arxiv.org/abs/2411.09915 (accessed 9 November 2026).
  3. Raissi M., Perdikaris P., Karniadakis G.E., "Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations", Journal of Computational Physics 378 (2019) 686-707.
  4. Wang S., Teng Y., Perdikaris P., "Understanding and mitigating gradient flow pathologies in physics-informed neural networks", SIAM Journal on Scientific Computing 43 (2021) A3055-A3081.
  5. Xue K. et al., "Temperature prediction of lithium-ion batteries based on physics-informed neural networks", Chinese Journal of Engineering, DOI 10.13374/j.issn2095-9389.2025.07.07.004 (2025).
  6. "Physics-informed neural networks for the fully coupled electrochemical-thermal pseudo-2D model of lithium-ion batteries", Elsevier (2026).
  7. "Towards Intelligent Systems for Battery Management: A Five-Tier Digital Twin Architecture", arXiv:2509.02366 (2025).
  8. Takamoto M. et al., "PDEBench: An Extensive Benchmark for Scientific Machine Learning", NeurIPS 2022 Datasets and Benchmarks Track.
  9. DClaw Research, "Physics-Informed Neural Networks in 2026: A Full-Stack Guide", site guide (9 November 2026).
  10. DClaw Research, "The Cost Boundary of Classical Solvers: When Physics-AI Surrogates Actually Pay Off in 2026", site guide (9 November 2026).
Tags: