⚡ electrical engineering

CFD-informed machine learning for surface-temperature mapping of side-cooled 4680 lithium-ion cells under thermal uncertainty

This study develops a computationally efficient machine learning surrogate, trained on 75 CFD simulations of a side-cooled 4680 lithium-ion cell under thermal uncertainty, to accurately reconstruct its full surface temperature field from limited sensor data, demonstrating that CNN-LSTM and SVM models achieve high predictive accuracy (R² ≈ 0.98) with axial position identified as the dominant temperature predictor.

Original authors: Aydin Ulus, Ihsan Uluocak, Wajih Nazim Kamal, Kyosung Choo, Young Jik Youn

Published 2026-11-04

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Aydin Ulus, Ihsan Uluocak, Wajih Nazim Kamal, Kyosung Choo, Young Jik Youn

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Inside the electric vehicles that power our modern world, the heart of the machine is a battery pack made of thousands of cylindrical cells. Among these, a specific large format known as the 4680 has become a standard for high-performance vehicles, offering more energy in a single unit than older designs. However, this size brings a hidden challenge: heat. As electricity flows through the battery, it generates warmth, and if that heat cannot escape evenly, the cell can degrade faster or even become unsafe. The problem is that the hottest part of the battery often lies deep inside, hidden beneath the metal casing, while sensors can only measure the temperature on the outside surface. Because the internal structure of the battery is complex and the way heat moves through it is not perfectly known, engineers struggle to predict exactly how hot the core gets just by looking at the skin.

To solve this puzzle, a team of researchers turned to a combination of computer simulation and artificial intelligence. They focused on a single 4680 cell that is cooled along its side, a common design in electric vehicles. The researchers knew that two specific factors create uncertainty in how well the battery cools down: the ability of the internal spiral layers to conduct heat outward, and the quality of the contact between the battery casing and the cooling hardware. Because these values can vary from one battery to another, a simple measurement of the surface temperature is often not enough to know the full thermal story. The team set out to build a digital tool that could take a few surface temperature readings and, accounting for these unknowns, reconstruct the surface temperature field of the battery's outer shell within the specific range of conditions they tested.

The researchers began by creating a highly detailed three-dimensional model of the battery cell inside a computer. This model acted as a virtual laboratory where they could run experiments that would be difficult or impossible to perform on real hardware. They simulated the flow of heat through the cell under steady conditions, with a fixed amount of energy being generated inside and a constant cooling temperature applied to the side. To test how the system reacts to uncertainty, they ran seventy-five different simulations. In each run, they slightly changed the internal heat-conducting ability of the battery's core and the thickness of the gap between the battery and the cooler. For every single simulation, they recorded the temperature at twenty-five specific points on the outside of the metal casing. This process generated a massive library of data, containing nearly two thousand distinct temperature snapshots that linked specific internal properties to specific surface readings.

With this library in hand, the team trained three different types of machine learning algorithms to act as a fast, intelligent translator. These algorithms were tasked with learning the relationship between the input variables—the location on the battery, the internal heat conductivity, and the contact gap—and the resulting surface temperature. The goal was to create a "surrogate" model, a lightweight software tool that could predict the temperature field almost instantly, without needing to run the heavy, time-consuming computer simulation every time. The researchers tested these tools by hiding a portion of their data and seeing if the algorithms could correctly guess the temperatures for the points they had not seen before.

The results showed that two of the the machine learning approaches performed with remarkable precision. One method, which combined a neural network designed to recognize patterns with a system that remembers sequences, and another method based on finding the best mathematical fit for the data, both achieved a coefficient of determination (R²) of nearly 0.98. Their predictions were off by less than a quarter of a degree Celsius on average. This level of accuracy means the tool can reliably recreate the temperature map of the battery surface across the range of conditions they tested. In contrast, a third, more standard type of neural network struggled to capture the subtle details, producing errors that were significantly larger and less consistent.

The study also revealed which factors mattered most when predicting the temperature. The most powerful predictor was simply the vertical position on the battery; the temperature changes significantly from the top to the bottom due to the way the cooling contact is arranged. The horizontal position around the cylinder was the second most important factor. Surprisingly, within the specific range of conditions the team tested, the exact value of the internal heat conductivity and the contact gap had a much smaller influence on the surface temperature than the physical location of the sensor. This suggests that for this specific cooling setup, knowing where you are on the battery tells you more about the temperature than the slight variations in the internal materials.

Using their best-performing model, the researchers demonstrated how the tool could visualize the temperature field for a specific set of conditions that lay inside their tested range but had not been simulated as a discrete data point. By feeding the model a new combination of internal properties, it generated a smooth, continuous map of the battery's surface temperature, showing exactly where the heat would concentrate and where it would dissipate. This visualization confirmed that the model could interpolate, or fill in the gaps, between the specific data points it had learned from, creating a coherent picture of the thermal behavior for that specific parameter combination.

The work provides a practical, computationally inexpensive way to understand the thermal state of these large batteries. By combining a validated physics model with machine learning, the researchers created a system that can quickly estimate the surface temperature distribution even when the internal properties are not perfectly known, but only within the steady-state domain they investigated. While the current tool is limited to steady-state conditions and does not yet account for rapid changes in power or aging, it offers a solid foundation for future thermal management systems. It proves that with the right data and the right algorithms, it is possible to see the invisible heat patterns inside a battery just by looking at its surface, a capability that is essential for keeping electric vehicles safe and efficient.

Technical Summary: CFD-Informed Machine Learning for Surface-Temperature Mapping of Side-Cooled 4680 Lithium-Ion Cells

Problem Statement
Large-format 4680 lithium-ion cells present significant challenges for thermal monitoring. Surface temperature measurements alone fail to fully characterize the internal thermal state, particularly given the substantial radial distance heat must travel from the jelly roll to the can surface. Furthermore, critical heat-transfer properties, such as radial jelly-roll conductivity and cell-to-cooler interface resistance, remain uncertain and variable. While existing literature has established the importance of cell architecture and cooling topologies, most analyses rely on fixed or calibrated thermal properties. Conversely, data-driven methods for internal temperature estimation often focus on specific parameterizations or operating conditions without adequately addressing simultaneous variations in thermal properties. This study addresses the gap in reconstructing the external can-surface temperature field of a side-cooled 4680 cell when radial jelly-roll conductivity and interface resistance vary simultaneously.

Methodology
The research employs a hybrid framework combining a validated three-dimensional steady-state Computational Fluid Dynamics (CFD) model with machine learning (ML) surrogates.

CFD Modeling: A 3D steady thermal model of a side-cooled Tesla 4680 Gen2 cell was developed in ANSYS Fluent. The geometry follows an experimentally validated configuration where an aluminum side cooler covers a 60° circumferential sector over a 70 mm axial height. The jelly roll is modeled as an orthotropic solid with fixed axial/circumferential conductivity ( $k_{\theta} = k_z = 23.1$ W m⁻¹ K⁻¹) and a variable radial conductivity ( $k_r$ ). An equivalent thin-wall thermal resistance represents the interface between the cell can and the cooler.
Experimental Design: A structured dataset was generated by varying two parameters: radial jelly-roll conductivity ( $k_r$ ) across 15 levels (0.90–2.00 W m⁻¹ K⁻¹) and equivalent interface thickness ( $t_{int}$ ) across 5 levels (0.40–0.60 mm). All other conditions, including a fixed 6.0 W heat input and a 25°C cooler boundary, were held constant. This resulted in 75 independent steady-state CFD cases.
Data Generation: From each CFD case, temperatures were extracted at 25 fixed locations on the outer can surface (a 5×5 grid of axial and circumferential positions), yielding a total of 1,875 spatial samples.
Machine Learning Framework: Three regression models were trained using point-wise data partitioning (five-fold cross-validation) and z-score normalization:
1. CNN-LSTM: A hybrid model treating the 25 surface locations as a sequence.
2. Support Vector Regression (SVM): A kernel-based non-neural comparator.
3. Multilayer Perceptron (MLP): A fully connected neural network baseline.
  Hyperparameters were optimized via Bayesian optimization.

Key Results

CFD Validation: The baseline CFD model reproduced three published validation temperatures with a maximum absolute difference of 0.196°C. Grid convergence analysis confirmed that the difference between the two finest meshes was less than 0.01%.
Model Performance: On the held-out test subset, the CNN-LSTM and SVM models demonstrated nearly identical high accuracy:
- CNN-LSTM: $R^2 = 0.981$ , RMSE = 0.238°C.
- SVM: $R^2 = 0.980$ , RMSE = 0.245°C.
- MLP: Performed significantly worse with $R^2 = 0.817$ and RMSE = 0.685°C, exhibiting broader error scatter and larger outliers.
Feature Importance: Perturbation analysis identified the axial position ( $z$ ) as the dominant predictor of surface temperature (increasing RMSE by ~1.8°C when perturbed), followed by the circumferential position ( $x$ ) and radial conductivity ( $k_r$ ). The influence of the interface resistance ( $t_{int}$ ) was comparatively small within the tested range.
Surrogate Reconstruction: The trained surrogate successfully reconstructed the can-surface temperature field for an interpolated parameter combination ( $k_r = 1.35$ W m⁻¹ K⁻¹, $t_{int} = 0.52$ mm), visualizing smooth axial and circumferential gradients consistent with the physical constraints of side cooling.

Significance and Claims
The paper claims that the developed CFD-informed ML framework provides a computationally inexpensive basis for thermal sensitivity analysis and surface temperature reconstruction within the investigated steady-state parameter range. The study demonstrates that machine learning surrogates can effectively learn the complex mapping between spatial coordinates and thermal properties to predict surface temperature fields.

However, the authors maintain a modest scope regarding the generalizability of these results. They explicitly state that the reported performance metrics quantify the interpolation of held-out spatial samples within the simulated design space, rather than independent generalization to completely unseen thermal operators or dynamic operating conditions. The models are restricted to steady-state behavior with fixed heat loads and cooling boundaries; they do not account for transient storage, aging effects, or thermal runaway. Consequently, the authors conclude that further case-wise validation is required before extending these models to unseen thermal operators or different operating conditions. The primary contribution is the demonstration of a surrogate modeling approach that captures the coupled response of surface temperature to variations in radial conductivity and interface resistance for side-cooled 4680 cells.

Technical Summary: CFD-Informed Machine Learning for Surface-Temperature Mapping of Side-Cooled 4680 Lithium-Ion Cells

More like this