Paper deep dive
AI-Driven Surrogate Models for Predicting Electrode-Scale Discharge Behavior in Lithium-Ion Batteries
Mengda Xing, Jean-Marie Lagniez, Alejandro Franco
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 91%
Last extracted: 7/28/2026, 3:33:26 AM
Summary
The paper introduces a deep learning surrogate pipeline based on the Swin3D Transformer to predict spatiotemporal discharge dynamics in lithium-ion batteries. The method integrates Gaussian Positional Encoding (GPE) for spatial feature representation and a Temporal Encoding module for time-series evolution. Validated on the Electrochemical Simulation (ES) dataset, the approach significantly outperforms point cloud baselines (PointNet, DGCNN, Point Transformer) in accuracy and reduces computational overhead from hours to milliseconds, enabling high-throughput battery design.
Entities (9)
Relation Signals (8)
Swin3D Transformer → uses → Gaussian Positional Encoding
confidence 95% · Our approach integrates... Gaussian Positional Encoding (GPE)... Swin3D Transformer
Swin3D Transformer → validateson → Electrochemical Simulation Dataset
confidence 95% · Experimental validation on an Electrochemical Simulation (ES) dataset
Swin3D Transformer → uses → Temporal Encoding Module
confidence 92% · integrates... a specialized Temporal Encoding module... Swin3D Transformer
Swin3D Transformer → outperforms → Point Transformer
confidence 90% · significantly outperforms state-of-the-art point cloud baselines... Point Transformer
Swin3D Transformer → outperforms → PointNet
confidence 90% · significantly outperforms state-of-the-art point cloud baselines... PointNet
Swin3D Transformer → outperforms → DGCNN
confidence 90% · significantly outperforms state-of-the-art point cloud baselines... DGCNN
Swin3D Transformer → reducescomputationtimefor → Lithium-Ion Battery
confidence 90% · reduces the computational overhead by orders of magnitude... for high-throughput battery design
Electrochemical Simulation Dataset → generatedby → COMSOL Multiphysics
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Physics-based simulations are essential for understanding the electrode-scale discharge behavior of lithium-ion batteries (LIBs) but suffer from prohibitive computational costs. To address this, we introduce a novel deep learning surrogate pipeline based on the Swin3D Transformer to predict spatiotemporal discharge dynamics directly from volumetric data. Our approach integrates two key innovations: Gaussian Positional Encoding (GPE), which enhances spatial feature representation by adapting to the complex geometry of electrode microstructures, and a specialized Temporal Encoding module to capture non-linear timeseries evolution. Experimental validation on an Electrochemical Simulation (ES) dataset demonstrates that our pipeline significantly outperforms state-of-the-art point cloud baselines in prediction accuracy. Furthermore, the proposed method reduces the computational overhead by orders of magnitude, providing a scalable and efficient framework for high-throughput battery design and optimization.
Tags
Links
- Source: https://arxiv.org/abs/2607.20577v2
- Canonical: https://arxiv.org/abs/2607.20577v2
Trouble viewing inline? Open PDF directly →
Full Text
27,683 characters extracted from source content.
Expand or collapse full text
AI-Driven Surrogate Models for Predicting Electrode-Scale Discharge Behavior in Lithium-Ion Batteries Mengda Xing 1[0000−0001−9838−4687] , Jean-Marie Lagniez 1[0000−0002−6557−4115] , and Alejandro A. Franco 2[0000−0001−7362−7849] 1 CRIL, UMR 8188, Université d’Artois, Rue Jean Souvraz SP 18, F-62307 Lens Cedex, France xing,lagniez@cril.fr 2 LRCS, UMR 7314, Université de Picardie Jules Verne, 15 rue Baudelocque, 80039 Amiens Cedex, France alejandro.franco@u-picardie.fr Abstract. Physics-based simulations are essential for understanding the electrode-scale discharge behavior of lithium-ion batteries (LIBs) but suf- fer from prohibitive computational costs. To address this, we introduce a novel deep learning surrogate pipeline based on the Swin3D Trans- former to predict spatiotemporal discharge dynamics directly from vol- umetric data. Our approach integrates two key innovations: Gaussian Positional Encoding (GPE), which enhances spatial feature representa- tion by adapting to the complex geometry of electrode microstructures, and a specialized Temporal Encoding module to capture non-linear time- series evolution. Experimental validation on an Electrochemical Simula- tion (ES) dataset demonstrates that our pipeline significantly outper- forms state-of-the-art point cloud baselines in prediction accuracy. Fur- thermore, the proposed method reduces the computational overhead by orders of magnitude, providing a scalable and efficient framework for high-throughput battery design and optimization. Keywords: Lithium-Ion Battery· Surrogate Modeling· Swin3D Trans- former· Spatiotemporal Forecasting· Scientific Machine Learning 1 Introduction Physics-based simulations using Finite Element Methods (FEM) are essential for analyzing Lithium-Ion Battery (LIB) microstructures but suffer from prohibitive computational costs, often requiring hours per discharge cycle [2]. While Ma- chine Learning (ML) surrogates offer acceleration, integrating them into legacy simulation frameworks (e.g., COMSOL, custom Fortran/C++ solvers) presents significant software engineering challenges. These monolithic systems lack the modularity to ingest modern AI components, making manual integration error- prone and hindering the adoption of "AI for Science" paradigms. arXiv:2607.20577v2 [cs.LG] 24 Jul 2026 2M. Xing et al. Fig. 1. Transition from (a) traditional numerical simulation to (b) the proposed AI- augmented surrogate modeling framework. To address this, we propose an automated framework that refactors rigid physics-based pipelines into hybrid, AI-augmented architectures. As illustrated in Fig. 1, our approach systematically replaces computationally intensive numer- ical solvers with deep learning surrogates. By treating AI adoption as a software refactoring process, we enable domain experts to evolve legacy systems with minimal manual intervention. Technically, we employ Swin3D Transformers enhanced with Gaussian Po- sitional Encoding to learn spatiotemporal dynamics directly from 3D unstruc- tured mesh data. This pipeline automates the end-to-end workflow from data extraction to model deployment ensuring physical consistency while achieving orders-of-magnitude acceleration. The main contributions of this work are: – Automated Refactoring Framework: We introduce a pipeline that mod- ernizes legacy battery simulation software by systematically integrating AI surrogates, improving both maintainability and efficiency. – Swin3D Surrogate Model: We propose a specialized deep learning ar- chitecture that directly processes volumetric mesh data, serving as a high- fidelity replacement for traditional solvers. – Efficiency Breakthrough: Validated on a large-scale Electrochemical Sim- ulation (ES) dataset, our approach reduces simulation time from hours to milliseconds without compromising accuracy. 2 Methodology As illustrated in Fig. 2, our pipeline predicts the spatiotemporal discharge be- havior of LIB cathodes. The framework consists of three stages: (1) converting AI-Driven Surrogate Models for LIB Discharge Prediction3 Electrode Parameters 3D Voxelization Electrode Geometry G(V, E) coord (n, 3) Surrogate Model n*c n*c + Temporal Decoding Gaussian Positional Encoding k Parameter Encoding Time n*c Fig. 2. Overview of the proposed point cloud forecasting pipeline. The system takes electrode geometry (red) as input, processes it through voxelization and a Swin3D- based surrogate model, and predicts the spatiotemporal discharge behavior (blue-to- green). unstructured mesh data into sparse voxel grids; (2) extracting geometric fea- tures using Gaussian Positional Encoding (GPE) and a Swin3D backbone; and (3) decoding temporal dynamics to generate time-series predictions. 2.1 Data Representation The input data is derived from physics-based simulations of electrode microstruc- tures [3]. We treat the finite element mesh nodes as a point cloud P =x i N i=1 ∈ R N×3 . To handle the high density of simulation points (approx. 10 5 points per sample) efficiently, we employ 3D Voxelization. This process quantizes the con- tinuous coordinates into a sparse voxel grid, reducing computational complexity while preserving the macroscopic geometric topology required for the Swin3D backbone. 2.2 Network Architecture Gaussian Positional Encoding (GPE). To compensate for the quantization loss in voxelization and capture sub-voxel geometry, we introduce GPE. Unlike static sine-cosine encodings, GPE utilizes learnable Gaussian kernels to map coordinates into high-dimensional feature spaces, effectively modeling complex non-Euclidean structures [1]. As defined in Eq. 1, for a pointx ∈R 3 and a kernel with centerμ and bandwidth σ: G(x,μ,σ) = exp − ∥x−μ∥ 2 2σ 2 (1) Bothμ and σ are learnable parameters. We stack outputs from K kernels to form the encoding vector GPE(x)∈R K . Surrogate Model & Temporal Decoding. We utilize the Swin3D Trans- former (Fig. 4) as the backbone. By employing shifted-window Multi-head Self- Attention (MSA), Swin3D efficiently captures both local fine-grained details and global semantic contexts within the voxel grid. To predict the time-series evolu- tion (24 time steps), we append a Temporal Encoding module, implemented as a Multi-Layer Perceptron (MLP), to the backbone’s output. 4M. Xing et al. 0.0 0.2 0.4 0.6 0.8 1.0 X 0.0 0.2 0.4 0.6 0.8 1.0 Y 0.0 0.2 0.4 0.6 0.8 1.0 Z Gaussian Positional Encoding Visualization Point Cloud Gaussian Centers Fig. 3. Gaussian Positional Encoding. Blue points: input coordinates; Red points: learnable Gaussian kernel centers. LN MSA MLP + LN LN Shift-MSA MLP + LN Two successive Swin3D blocks + + + + Data flow Encoding Layer Swin3d Block*N Downsample Decoding Layer Swin3d Block*N Upsample Fig. 4. Left: The Swin3D backbone architecture. Right: Details of two successive Swin3D Blocks with shifted windows. Optimization. The network is trained in an end-to-end manner using the Mean Squared Error (MSE) loss: L = 1 n P (Y gt − ˆ Y pred ) 2 , aiming to minimize the discrepancy between the predicted electrochemical fields and the physics-based ground truth. 3 Related Work We review existing literature on 3D deep learning, categorizing methods into two primary streams: point-based architectures and transformer-based architectures. Point-based Methods. PointNet [5] pioneered the direct processing of un- ordered point clouds using deep learning. It employs a shared Multi-Layer Per- ceptron (MLP) to process each point independently, followed by a symmetric max-pooling function to aggregate global features. While efficient, PointNet AI-Driven Surrogate Models for LIB Discharge Prediction5 struggles to capture local geometric structures due to its lack of local context awareness. To address this, PointNet++ [6] introduced a hierarchical learning mechanism that recursively applies PointNet to local neighborhoods, enabling the extraction of fine-grained geometric details at multiple scales. Further im- proving local feature learning, DGCNN [8] proposed the EdgeConv operation, which constructs dynamic k-nearest neighbor (k-N) graphs in the feature space. This allows the network to explicitly model the relationship between a point and its neighbors, learning geometry-aware embeddings that capture the topological structure of the data. Transformer-based Methods. Inspired by the success of Transformers in nat- ural language processing and 2D vision, recent works have adapted self-attention mechanisms to 3D data. Point Transformer [13] introduces a vector self-attention block specifically designed for point clouds. By leveraging positional encoding and attention within local neighborhoods, it effectively captures both geomet- ric structures and global dependencies, outperforming traditional MLP-based methods on semantic segmentation tasks. Building on this, Point Transformer V2 (PTv2) [9] incorporates grouping vector attention and partition-based pool- ing to further enhance computational efficiency and feature integration. Moving beyond supervised learning, Point-BERT [12] introduces a masked point mod- eling paradigm, adapting the BERT pre-training framework to learning generic 3D representations via masked auto-encoding. Most relevant to our work is the Swin3D Transformer [11]. Building upon the hierarchical design of the Swin Transformer [4], Swin3D extends the shifted window attention mechanism to the 3D domain. Unlike point-based methods that operate on unstructured coordinates, Swin3D processes voxelized volumet- ric data, making it particularly effective for tasks requiring structured spatial understanding, such as medical image analysis and large-scale scene understand- ing [10, 7]. This ability to handle structured 3D grids makes it an ideal backbone for predicting dense physical fields in battery simulations. 4 Experiments 4.1 Experimental Setup The surrogate models were implemented using PyTorch 2.3.0 on a workstation running Ubuntu 22.04. The hardware configuration includes an NVIDIA RTX A4500 GPU (20 GB VRAM) and 64 GB of RAM. All training and inference processes were GPU-accelerated. The physics-based ground truth data was gen- erated using COMSOL Multiphysics, a standard finite element analysis solver for electrochemical modeling. 4.2 Experimental Settings Baselines. To evaluate the effectiveness of our proposed method, we compare Swin3D against representative state-of-the-art point cloud learning architectures. 6M. Xing et al. For point-based methods, we select PointNet [5], which processes points inde- pendently, and DGCNN [8], which utilizes dynamic graph convolutions. For transformer-based methods, we compare against Point Transformer (PT) [13] and Point Transformer V2 (PTv2) [9], which apply self-attention mechanisms to point sets. All baselines were retrained on the ES dataset using identical protocols to ensure a fair comparison. Model Variants. We analyze the impact of positional encoding and model capacity through various configurations: – Positional Encodings: We compare our proposed Gaussian Positional Encoding (GPE) against Sine-Cosine (SIN) encoding and raw coordinates (XYZ). Unlike fixed SIN functions, GPE employs learnable Gaussian ker- nels that dynamically adapt to the spatial distribution of the electrode mi- crostructures, thereby capturing intricate geometric details more effectively. – Model Capacity: We evaluate four scales of the Swin3D architecture: Tiny (T), Small (S), Medium (M), and Large (L). The channel configurations for the five stages are defined as follows: • Swin3D-T: C =32, 64, 128, 256, 256 • Swin3D-S: C =48, 96, 192, 384, 384 • Swin3D-M: C =64, 128, 256, 512, 512 • Swin3D-L: C =80, 160, 320, 640, 640 Smaller variants (T, S) prioritize computational efficiency, while larger variants (M, L) offer higher theoretical capacity. Dataset. The summary of the dataset is shown in Table 1. We validate our pipeline on the Electrochemical Simulation (ES) dataset, which models the dis- charge process of an NMC111 cathode. The dataset contains high-fidelity 3D fields of lithium concentration and electric potential within the active material and carbon binder domains. 4.3 Evaluation Metrics We employ both pointwise error maps for visual assessment and aggregated statistical metrics for numerical comparison. Visual Metrics. To visualize the spatial distribution of errors, we define the Pointwise Error (PE) as the difference between the prediction ˆ Y and the ground truth Y . The Mean Absolute Error (MAE) aggregates this over all n points: PE = ˆ Y − Y,MAE = 1 n n X i=1 | ˆ Y i − Y i |(2) Note that in the results section, we visualize PE to identify regions of under- or over-estimation. AI-Driven Surrogate Models for LIB Discharge Prediction7 Table 1. Summary of the Electrochemical Simulation (ES) Dataset. The dataset covers varying electrode designs and operational conditions, providing dense 3D concentration fields over time. AttributeSpecification Geometry & Scale Dimensions (W × D× H) 34.5× 34.5× 123 (unitless) Number of Points9.4× 10 5 ∼ 2.4× 10 6 per sample Number of Cells (Tetrahedra) 5.5× 10 5 ∼ 7.4× 10 5 per sample Simulation Parameters Active Material (AM)87%, 93%, 95% (wt) Calendering Degree (Cal)0%, 10%, 20% Discharge Rate (C-rate)0.5C, 1C, 2C Data Structure Time Steps24 steps (t = 0 to t = 23) Target VariableLi-Ion Concentration (C) Total Samples18 (16 Training, 2 Testing) Numerical Metrics. We use three standard metrics to quantify overall per- formance: Root Mean Squared Error (RMSE) for error magnitude, Symmetric Mean Absolute Percentage Error (SMAPE) for relative accuracy, and the Coef- ficient of Determination (R 2 ) for explanatory power: RMSE = v u u t 1 n n X i=1 Y i − ˆ Y i 2 (3) SMAPE = 100% n n X i=1 | ˆ Y i − Y i | (| ˆ Y i | +|Y i |)/2 (4) R 2 = 1− P n i=1 (Y i − ˆ Y i ) 2 P n i=1 (Y i − ̄ Y ) 2 (5) where ̄ Y represents the mean of the observed data. Higher R 2 and lower RMSE/SMAPE indicate better performance. 5 Results In this section, we evaluate the performance of our proposed method on the electrochemical simulation dataset. We present a comprehensive analysis includ- ing visual quality assessment, quantitative numerical comparison, computational efficiency analysis, and ablation studies on the key components: Gaussian Posi- tional Encoding (GPE) and Temporal Encoding. 8M. Xing et al. T=6 DGCNN AvRE=0.3754 POINTNET AvRE=1.0779 PT_XYZ AvRE=0.8681 PTV2_XYZ AvRE=1.0928 Swin3D_T_GPE AvRE=0.5251Swin3D_S_GPE AvRE=0.1974Swin3D_M_GPE AvRE=0.9892Swin3D_L_GPE AvRE=1.0134 T=12 DGCNN AvRE=0.2315 POINTNET AvRE=0.9344 PT_XYZ AvRE=0.7718 PTV2_XYZ AvRE=0.9510 Swin3D_T_GPE AvRE=0.4162Swin3D_S_GPE AvRE=0.0914Swin3D_M_GPE AvRE=0.8723Swin3D_L_GPE AvRE=0.9007 T=24 DGCNN AvRE=0.2061 POINTNET AvRE=0.7164 PT_XYZ AvRE=0.6077 PTV2_XYZ AvRE=0.7147 Swin3D_T_GPE AvRE=0.3055Swin3D_S_GPE AvRE=0.1237Swin3D_M_GPE AvRE=0.5798Swin3D_L_GPE AvRE=0.6547 2 1 0 1 2 2 1 0 1 2 2 1 0 1 2 2 1 0 1 2 2 1 0 1 2 2 1 0 1 2 2 1 0 1 2 2 1 0 1 2 2 1 0 1 2 2 1 0 1 2 2 1 0 1 2 2 1 0 1 2 2 1 0 1 2 2 1 0 1 2 2 1 0 1 2 2 1 0 1 2 2 1 0 1 2 2 1 0 1 2 2 1 0 1 2 2 1 0 1 2 2 1 0 1 2 2 1 0 1 2 2 1 0 1 2 2 1 0 1 2 Fig. 5. Visual comparison of the Relative Error (RE) in lithium-ion concentration predictions. The results are visualized as 2D slices extracted from the 3D volumetric predictions. Each column corresponds to a different model, and each row represents a specific time step (t = 6, 12, 24). Ideally, the error map should be gray (indicating zero error). 5.1 Visual Comparison We first assess the visual fidelity of the lithium-ion concentration predictions. Fig. 5 illustrates the spatial distribution of the Relative Error (RE) across dif- ferent time steps (t = 6, 12, 24). The visual results demonstrate that the proposed Swin3D architecture signif- icantly outperforms other surrogate models. As shown in the error maps (where red and blue indicate large positive and negative errors, respectively), Swin3D maintains a consistently low error rate across all time steps. Quantitatively, the Swin3D_S_GPE variant achieves the lowest Average Relative Error (AvRE) of 0.1974, 0.0914, and 0.1237 at time steps t = 6, 12, and 24, respectively. In con- trast, competing models exhibit significantly higher error variance, underscoring the effectiveness of our method in capturing intricate spatial-temporal features. We further investigate the model’s capability to predict discharge behaviors under varying electrode parameters. Fig. 6 displays the discharge curves for 12 distinct parameter sets. As shown in Fig. 6, the majority of parameter combinations yield an AvRE below 3%. Outlier cases (95_dried_0.5, 95_cal10_2) exhibit higher errors (∼9– 10%), attributable to a distribution shift where the discharge potential peaks at 3.5 V rather than the typical 3.6 V, a limitation to be addressed in future work. 5.2 Numerical Comparison Table 2 presents the quantitative comparison of our method against state-of- the-art point cloud learning baselines, including PointNet [5], DGCNN [8], Point Transformer [13], and Point Transformer V2 [9]. The Swin3D_S_GPE model achieves the best performance across all metrics, with an RMSE of 0.0734, SMAPE of 55.43, and R 2 of 0.7946. Interestingly, AI-Driven Surrogate Models for LIB Discharge Prediction9 2500030000350004000045000 3.6 3.7 3.8 3.9 4.0 4.1 93_cal20_2 AvRE = 0.87% 25000300003500040000 3.5 3.6 3.7 3.8 3.9 4.0 4.1 95_cal20_1 AvRE = 3.96% 25000300003500040000 3.7 3.8 3.9 4.0 4.1 95_cal10_0.5 AvRE = 1.27% 25000300003500040000 3.7 3.8 3.9 4.0 4.1 93_cal20_1 AvRE = 2.09% True Swin3d(ours) 2500030000350004000045000 3.5 3.6 3.7 3.8 3.9 4.0 4.1 95_dried_1 AvRE = 5.18% 25000300003500040000 3.7 3.8 3.9 4.0 4.1 93_cal20_0.5 AvRE = 2.84% 2500030000350004000045000 3.6 3.7 3.8 3.9 4.0 4.1 87_cal0_2 AvRE = 1.85% 25000300003500040000 3.7 3.8 3.9 4.0 4.1 87_cal0_0.5 AvRE = 2.55% 2500030000350004000045000 3.7 3.8 3.9 4.0 4.1 95_dried_0.5 AvRE = 1.75% 25000300003500040000 3.4 3.6 3.8 4.0 95_dried_2 AvRE = 9.9% 2500030000350004000045000 3.5 3.6 3.7 3.8 3.9 4.0 4.1 95_cal10_1 AvRE = 4.66% 25000300003500040000 3.4 3.6 3.8 4.0 95_cal10_2 AvRE = 9.28% Concentration Li-ion AM Potential in Solid Phase Fig. 6. Discharge curves for 12 different sets of electrode parameters. the "Small" (S) variant of our architecture outperforms the "Large" (L) and "Medium" (M) variants. This counter-intuitive result can be attributed to the regularization effect of the smaller model capacity. Given the limited size of the electrochemical simulation dataset, the larger models are prone to overfitting, whereas the smaller architecture strikes an optimal balance between feature ex- pressiveness and generalization. 5.3 Computational Efficiency One of the primary motivations for using a deep learning surrogate model is to accelerate the time-consuming physics-based simulations. Table 3 compares the computational costs. The physics-based simulation takes approximately 13 hours (46,800 seconds) to compute a single discharge process. In stark contrast, our Swin3D_S_GPE model requires only 0.1487 seconds for inference. This represents a speedup of over magnitudes of 10 5 , enabling real-time prediction and high-throughput material screening that were previously computationally infeasible. 5.4 Ablation Studies Effectiveness of Gaussian Positional Encoding (GPE). We evaluate the impact of different positional encoding strategies: standard Cartesian coordi- 10M. Xing et al. Table 2. Numerical comparison of different surrogate models on the Concentration of Electrochemical Simulation dataset. RMSE: Root Mean Square Error; SMAPE: Symmetric Mean Absolute Percentage Error. Best results are highlighted in bold. MethodsRMSE ↓ SMAPE ↓R 2 ↑ PointNet (CVPR17)0.4872124.5293-0.1633 DGCNN (SIGGRAPH19)0.240592.80540.1945 PT (ICCV21)0.2348115.00830.3425 PTV2 (NeurIPS22)0.4278120.8603-3.7154 Swin3D_L_GPE0.153586.09730.5703 Swin3D_M_GPE0.214092.49200.4009 Swin3D_S_GPE0.073455.4275 0.7946 Swin3D_T_GPE0.126984.70200.6448 Table 3. Runtime comparison between physics-based Electrochemical Simulation (ES) and AI-driven surrogate models. The ES runtime is for a single simulation instance. MethodTraining Time Inference Time (seconds/epoch) (seconds/sample) Physics-Based SimulationN/A≈ 46,800 (13 h) PointNet (CVPR17)1.050.0130 DGCNN (SIGGRAPH19)2.230.0279 PT (ICCV21)6.360.0805 PTV2 (NeurIPS22)12.710.1569 Swin3D_S_GPE (Ours)11.890.1487 nates (XYZ), Sinusoidal encoding (SIN), and our proposed Gaussian Positional Encoding (GPE). Table 4. Ablation study of positional encoding strategies. Comparisons are made across different model sizes (T, S, M, L). ModelEncoding RMSE ↓ SMAPE ↓R 2 ↑ Swin3D_SXYZ0.096469.77790.7300 Swin3D_SSIN0.095472.22040.7328 Swin3D_S GPE0.073455.4275 0.7946 Swin3D_TGPE0.126984.70200.6448 Swin3D_MGPE0.214092.49200.4009 Swin3D_LGPE0.153586.09730.5703 As shown in Table 4, GPE consistently outperforms XYZ and SIN encodings. For the best-performing ‘S‘ model, using GPE reduces the RMSE from 0.0964 (XYZ) to 0.0734. Fig. 7 visually confirms this, showing that models trained with GPE produce smoother and more accurate concentration fields. This suggests that GPE provides a more expressive and adaptive representation for the com- plex geometries found in electrode microstructures. AI-Driven Surrogate Models for LIB Discharge Prediction11 T=6 Swin3D_T_XYZ AvRE=0.5046Swin3D_T_GPE AvRE=0.5251Swin3D_T_SIN AvRE=0.3351Swin3D_S_XYZ AvRE=0.5502Swin3D_S_GPE AvRE=0.1974Swin3D_S_SIN AvRE=0.4847Swin3D_M_XYZ AvRE=1.3473Swin3D_M_GPE AvRE=0.9892Swin3D_M_SIN AvRE=0.9843Swin3D_L_XYZ AvRE=1.2235Swin3D_L_GPE AvRE=1.0134Swin3D_L_SIN AvRE=0.6771 T=12 Swin3D_T_XYZ AvRE=0.4028Swin3D_T_GPE AvRE=0.4162Swin3D_T_SIN AvRE=0.2681Swin3D_S_XYZ AvRE=0.4000Swin3D_S_GPE AvRE=0.0914Swin3D_S_SIN AvRE=0.4611Swin3D_M_XYZ AvRE=1.1849Swin3D_M_GPE AvRE=0.8723Swin3D_M_SIN AvRE=0.8488Swin3D_L_XYZ AvRE=1.0716Swin3D_L_GPE AvRE=0.9007Swin3D_L_SIN AvRE=0.6072 T=24 Swin3D_T_XYZ AvRE=0.3672Swin3D_T_GPE AvRE=0.3055Swin3D_T_SIN AvRE=0.2713Swin3D_S_XYZ AvRE=0.3808Swin3D_S_GPE AvRE=0.1237Swin3D_S_SIN AvRE=0.3282Swin3D_M_XYZ AvRE=0.7466Swin3D_M_GPE AvRE=0.5798Swin3D_M_SIN AvRE=0.5966Swin3D_L_XYZ AvRE=0.7156Swin3D_L_GPE AvRE=0.6547Swin3D_L_SIN AvRE=0.5119 2 1 0 1 2 2 1 0 1 2 2 1 0 1 2 2 1 0 1 2 2 1 0 1 2 2 1 0 1 2 2 1 0 1 2 2 1 0 1 2 2 1 0 1 2 2 1 0 1 2 2 1 0 1 2 2 1 0 1 2 2 1 0 1 2 2 1 0 1 2 2 1 0 1 2 2 1 0 1 2 2 1 0 1 2 2 1 0 1 2 2 1 0 1 2 2 1 0 1 2 2 1 0 1 2 2 1 0 1 2 2 1 0 1 2 2 1 0 1 2 2 1 0 1 2 2 1 0 1 2 2 1 0 1 2 2 1 0 1 2 2 1 0 1 2 2 1 0 1 2 2 1 0 1 2 2 1 0 1 2 2 1 0 1 2 2 1 0 1 2 2 1 0 1 2 2 1 0 1 2 Fig. 7. Visual comparison of prediction errors using different positional encodings (XYZ, SIN, GPE). The GPE variant shows reduced error magnitudes in the high- lighted regions. Configuration of Temporal Encoding. Finally, we optimized the Temporal Encoding module through a grid search. We explored variations in network depth (1-3 hidden layers), width (0.5C to 2C), and activation functions (ReLU vs. GeLU). The experimental results indicate that a configuration with 1 hidden layer, 2C neurons, and ReLU activation yields the best performance. This simple yet effective structure is sufficient to capture the temporal dependencies of the discharge process without introducing unnecessary complexity. 6 Conclusion In this paper, we presented a novel automated framework for integrating deep learning surrogate models into physics-based battery simulation workflows. By leveraging a Swin3D backbone enhanced with Gaussian Positional Encoding and a specialized Temporal Encoding module, our approach effectively captures the complex spatial-temporal dynamics of electrode discharge processes directly from volumetric mesh data. Empirical results on the Electrochemical Simulation (ES) dataset demon- strate that our model not only outperforms state-of-the-art point cloud baselines in prediction accuracy but also achieves a dramatic reduction in computational time-accelerating simulations from hours to milliseconds. This efficiency break- through paves the way for high-throughput material screening and real-time battery management, which were previously computationally prohibitive. Beyond the specific domain of battery modeling, the proposed pipeline offers a generalized solution for 3D point cloud forecasting tasks in scientific computing. Future work will focus on extending this framework to handle multi-physics coupling scenarios and exploring active learning strategies to further minimize the data dependence of the surrogate models. Acknowledgments. We thank the reviewers for their insightful feedback and con- structive suggestions. Disclosure of Interests. The authors have no competing interests to declare that are relevant to the content of this article. 12M. Xing et al. References 1. Chandra, M.A., Bedi, S.: Survey on svm and their application in image classifica- tion. International Journal of Information Technology 13(5), 1–11 (2021) 2. Duquesnoy, M., Liu, C., Dominguez, D.Z., Kumar, V., Ayerbe, E., Franco, A.A.: Machine learning-assisted multi-objective optimization of battery manufacturing from synthetic data generated by physics-based simulations. Energy Storage Ma- terials 56, 50–61 (2023) 3. Liu, C., Lombardo, T., Xu, J., Ngandjong, A.C., Franco, A.A.: An experimentally-validated 3d electrochemical model revealing electrode manufac- turing parameters’ effects on battery performance. Energy Storage Materials 54, 156–163 (2023). https://doi.org/https://doi.org/10.1016/j.ensm.2022.10.035, https://w.sciencedirect.com/science/article/pii/S2405829722005591 4. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: Proceedings of the IEEE/CVF international conference on computer vision. p. 10012–10022 (2021) 5. Qi, C.R., Su, H., Mo, K., Guibas, L.J.: Pointnet: Deep learning on point sets for 3d classification and segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition. p. 652–660 (2017) 6. Qi, C.R., Yi, L., Su, H., Guibas, L.J.: Pointnet++: Deep hierarchical feature learn- ing on point sets in a metric space. Advances in neural information processing systems 30 (2017) 7. Siddiqui, Y., Porzi, L., Buló, S.R., Müller, N., Nießner, M., Dai, A., Kontschieder, P.: Panoptic lifting for 3d scene understanding with neural fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. p. 9043–9052 (2023) 8. Wang, Y., Sun, Y., Liu, Z., Sarma, S.E., Bronstein, M.M., Solomon, J.M.: Dynamic graph cnn for learning on point clouds. ACM Transactions on Graphics (tog) 38(5), 1–12 (2019) 9. Wu, X., Lao, Y., Jiang, L., Liu, X., Zhao, H.: Point transformer v2: Grouped vector attention and partition-based pooling. Advances in Neural Information Processing Systems 35, 33330–33342 (2022) 10. Xue, R., Zhang, X., Soergel, U.: Urban classification based on top-view point cloud and sar image fusion with swin transformer. The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences 43, 559–564 (2022) 11. Yang, Y.Q., Guo, Y.X., Xiong, J.Y., Liu, Y., Pan, H., Wang, P.S., Tong, X., Guo, B.: Swin3d: A pretrained transformer backbone for 3d indoor scene understanding (2023) 12. Yu, X., Tang, L., Rao, Y., Huang, T., Zhou, J., Lu, J.: Point-bert: Pre-training 3d point cloud transformers with masked point modeling. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. p. 19313– 19322 (2022) 13. Zhao, H., Jiang, L., Jia, J., Torr, P.H., Koltun, V.: Point transformer. In: Proceed- ings of the IEEE/CVF international conference on computer vision. p. 16259– 16268 (2021)