Local-Global CNN (LGCNN) in Scientific Modeling
- The paper introduces LGCNN as a hybrid model integrating local CNN feature extraction with an explicit global mechanism (e.g., a numerical solver) to address nonlocal dependencies.
- It clearly delineates local processing for predicting intermediate fields and a global pathway for computing streamlines or enhancing features to capture extended spatial correlations.
- Empirical results demonstrate that LGCNN can achieve competitive accuracy with minimal training data and scale effectively to larger domains compared to fully data-driven models.
Searching arXiv for papers on Local-Global Convolutional Neural Networks and related local-global CNN formulations. Local-Global Convolutional Neural Network (LGCNN) denotes a class of architectures that couple locality-preserving convolutional learning with an explicit mechanism for representing or injecting nonlocal, scene-level, or domain-level structure. In the most specific usage available in the literature provided here, LGCNN is the name of a hybrid surrogate for groundwater heat transport that combines convolutional neural networks with a lightweight numerical transport solver (Pelzer et al., 8 Jul 2025). In a broader but less formal sense, closely related “local-global” CNN designs appear in image enhancement, geospatial forecasting, visual local features, and classification, where convolutional processing is supplemented by global encoders, attention-like modules, or position-dependent components rather than relying on local receptive fields alone (Kinoshita et al., 2019, Uselis et al., 2020, Miao et al., 2022, Rezvani et al., 2023).
1. Terminology and scope
The term “Local-Global Convolutional Neural Network” is not used uniformly across domains. One paper introduces LGCNN explicitly as a method name for a groundwater heat transport surrogate (Pelzer et al., 8 Jul 2025). Several other papers implement the same architectural principle without using that exact acronym: a CNN “considering both local and global features” for image enhancement (Kinoshita et al., 2019), “localized convolutional neural networks” that learn local features in addition to global ones for wind forecasting (Uselis et al., 2020), a local feature method with “global visual information enhancement” (Miao et al., 2022), and a two-stream “Global/Local Processing” model (Rezvani et al., 2023).
Across these formulations, the recurrent design objective is consistent: standard CNNs are strong local feature extractors but are limited when the task depends on long-range coupling, scene-level exposure context, holistic shape, or absolute spatial location. This suggests that LGCNN is best understood not as a single canonical layer type, but as an architectural pattern in which local convolution is retained while a distinct mechanism handles nonlocal dependencies, global priors, or location-specific structure.
A common misconception is that any model with a large receptive field is automatically “local-global.” The literature summarized here supports a narrower interpretation. The defining feature is not merely depth or dilation, but an explicit structural separation between a local pathway and a global pathway or prior. In some cases the global component is another neural branch; in others it is a numerical solver or a global physical model (Pelzer et al., 8 Jul 2025, Lu et al., 2024).
2. Core architectural principle
The most explicit LGCNN decomposition is given for groundwater heat transport. There, the architecture is a three-step pipeline: $\text{CNN}(p, k, i) = \Vec v,$
$\text{IVP}(i, \Vec v) = \vec s,$
$\text{CNN}(p, k, i, \Vec v, \Vec s) = T.$
Here is the hydraulic pressure field, the heterogeneous permeability field, a one-hot encoding of GWHP pump locations, $\Vec v = (v_x, v_y)$ the predicted velocity field, the streamline representation, and the temperature field (Pelzer et al., 8 Jul 2025). The first and third stages are local CNN mappings; the middle stage is the global mechanism.
This division of labor is the central LGCNN idea. CNNs are used where locality dominates, while a separate mechanism captures the nonlocal process that would otherwise be difficult for convolutions to infer robustly. In the groundwater setting, the paper states that the system is advection dominated, with , and that small perturbations in $\text{IVP}(i, \Vec v) = \vec s,$0 can produce large downstream changes in plume trajectories. The global stage therefore computes streamlines through the ordinary differential equation
$\text{IVP}(i, \Vec v) = \vec s,$1
using SciPy’s solve_ivp with an implicit fifth-order Runge–Kutta method and 10,000 time steps of one day (Pelzer et al., 8 Jul 2025).
Related papers instantiate the same principle differently. In image enhancement, a local encoder processes the original-resolution image, a global encoder processes a resized $\text{IVP}(i, \Vec v) = \vec s,$2 view of the full image, and a decoder fuses both streams (Kinoshita et al., 2019). In local feature extraction, a lightweight CNN encoder is augmented by a Global Enhancement Module based on a Non-local block with positional embedding (Miao et al., 2022). In classification, a pretrained CNN local stream is concatenated with a global stream called the Global Advantage Stream, which uses adaptive low-pass filtering and lightweight convolutional processing (Rezvani et al., 2023). These variants differ in mechanism, but not in the underlying commitment to explicit local-global factorization.
3. The canonical LGCNN formulation for groundwater heat transport
The groundwater formulation is the clearest instance in which the acronym LGCNN is the method’s proper name. The task is subsurface temperature prediction for open-loop groundwater heat pumps in a heterogeneous porous medium (Pelzer et al., 8 Jul 2025). The difficulty arises because heat plume evolution is not purely local: it depends on the global hydraulic pressure gradient and long-range transport along streamlines, while also remaining highly sensitive to the local permeability field.
The model operationalizes this by first predicting groundwater velocity from local spatial inputs, then tracing global transport paths numerically, and finally predicting the temperature field from the original inputs plus the intermediate physics-informed outputs. To make the streamline representation useful to the second CNN, paths are embedded into the grid as spatial maps with values that fade linearly from 1 to 0 along the path. The method also computes outer streamlines $\text{IVP}(i, \Vec v) = \vec s,$3 by perturbing each pump location by 10 cells orthogonal to the global flow direction $\text{IVP}(i, \Vec v) = \vec s,$4, yielding a streamline input
$\text{IVP}(i, \Vec v) = \vec s,$5
The architecture can therefore be summarized as
$\text{IVP}(i, \Vec v) = \vec s,$6
The paper describes this as Step 1: CNN $\text{IVP}(i, \Vec v) = \vec s,$7, Step 2: simplified solver $\text{IVP}(i, \Vec v) = \vec s,$8, and Step 3: CNN $\text{IVP}(i, \Vec v) = \vec s,$9 (Pelzer et al., 8 Jul 2025).
This formulation is notable because the global component is not an attention block or pooling operator, but an explicit numerical surrogate. A plausible implication is that the paper treats “global” not as a representational scale alone, but as a computational regime: the part of the problem dominated by long-range transport is handed to a solver, while the parts dominated by local mappings are left to CNNs.
4. Variants of local-global fusion in related CNN literature
The local-global idea has several recurring implementation patterns in adjacent work.
| Paper | Local component | Global component |
|---|---|---|
| "Few-Shot Learning by Explicit Physics Integration: An Application to Groundwater Heat Transport" (Pelzer et al., 8 Jul 2025) | CNN for $\text{CNN}(p, k, i, \Vec v, \Vec s) = T.$0 and CNN for $\text{CNN}(p, k, i, \Vec v, \Vec s) = T.$1 | IVP-based streamline surrogate |
| "Convolutional Neural Networks Considering Local and Global features for Image Enhancement" (Kinoshita et al., 2019) | Local encoder on original-resolution image | Global encoder on resized $\text{CNN}(p, k, i, \Vec v, \Vec s) = T.$2 image |
| "Real-time Local Feature with Global Visual Information Enhancement" (Miao et al., 2022) | CNN encoder and multi-scale local feature maps | Non-local block with positional embedding |
| "Global-Local Processing in Convolutional Neural Networks" (Rezvani et al., 2023) | Conventional pretrained CNN backbone | Global Advantage Stream with adaptive low-pass filtering |
| "Localized convolutional neural networks for geospatial wind forecasting" (Uselis et al., 2020) | Shared convolutions | Learnable inputs, local weights, and persistent localization |
In the image-enhancement architecture, the local encoder is “almost the same as the encoder part of U-Net,” while the global encoder receives a resized $\text{CNN}(p, k, i, \Vec v, \Vec s) = T.$3 version of the image and all layers have 64 filters (Kinoshita et al., 2019). The decoder combines features from both branches, with concatenated skip connections from the local encoder. The paper explicitly reports that removing the global encoder introduces block-like artifacts and distortions.
In the local-feature setting, the deepest feature map is enhanced by a Non-local operation with 2D absolute sinusoidal positional embedding (Miao et al., 2022). The enhanced feature is
$\text{CNN}(p, k, i, \Vec v, \Vec s) = T.$4
where
$\text{CNN}(p, k, i, \Vec v, \Vec s) = T.$5
This is a neural global-aggregation mechanism rather than the solver-based global stage of the groundwater LGCNN.
In geospatial forecasting, the local-global distinction is framed differently. Standard CNNs capture shared spatial laws, but the same pattern may behave differently at different locations. The proposed localization mechanisms therefore add learnable inputs or location-specific local weights to ordinary convolutions, preserving global convolutional sharing while permitting local deviations (Uselis et al., 2020). This suggests another interpretation of “global” within LGCNN-like models: not long-range transport, but shared structure across space.
5. Training regimes, data efficiency, and scaling behavior
A major theme in the explicit LGCNN paper is few-shot learning under scientific-data scarcity. The synthetic baseline dataset spans a domain size of $\text{CNN}(p, k, i, \Vec v, \Vec s) = T.$6 on a $\text{CNN}(p, k, i, \Vec v, \Vec s) = T.$7 grid, with permeability generated using Perlin noise (Pelzer et al., 8 Jul 2025). Two setups are reported: 3dp, consisting of three simulations with different $\text{CNN}(p, k, i, \Vec v, \Vec s) = T.$8 and $\text{CNN}(p, k, i, \Vec v, \Vec s) = T.$9, one datapoint each for train/val/test, plus an additional 4× larger simulation for scaling; and 101dp, with 101 datapoints split 73:18:10 for train:val:test, used for purely data-driven baselines.
The few-shot behavior is achieved through patch partitioning. The best synthetic configuration uses 20,736 overlapping patches of size 0 for Step 1 and 82,944 patches for Step 3 (Pelzer et al., 8 Jul 2025). The paper explicitly notes that the approach can work with as little as a single simulation run. On realistic data, the model is trained on a handful of cut-outs from borehole-derived permeability maps in the Munich region, using 4 simulations for training/validation and 1 larger simulation for testing scalability.
A defining property of the model is scaling to larger domains without retraining. Because the CNNs are local and the global transport is handled by the numerical streamline step, the method can be applied directly to bigger maps (Pelzer et al., 8 Jul 2025). On a synthetic domain 4× larger than the training domain, the isolated Step 3 MAE is about 0.02 °C and the full pipeline MAE about 0.04 °C. On the Munich-derived dataset, the scaling domain is twice the length of the training domain, and the method remains stable.
This contrasts with the image-enhancement local-global CNN, where the benefit of the global branch is not domain scaling but improved global consistency and avoidance of local-only distortions (Kinoshita et al., 2019). It also differs from the geospatial localization literature, where local-global augmentation is introduced to correct translation-invariance assumptions rather than to enable large-domain extrapolation (Uselis et al., 2020). These comparisons indicate that LGCNN is a general architectural motif, but the specific benefit depends on the problem’s nonlocal structure.
6. Empirical performance and comparative evidence
For the canonical LGCNN paper, the reported results are organized by stage and by comparison with data-driven baselines (Pelzer et al., 8 Jul 2025). On the synthetic 3dp test set, Step 1 velocity prediction yields 1: MAE 2 m/y, SSIM 3, and 4: MAE 5 m/y, SSIM 6. Isolated Step 3 temperature prediction on synthetic test gives 7, MAE 8, PAT 9, and SSIM 0. Using predicted velocities in the full pipeline increases the synthetic test MAE to 1, which the paper attributes to sensitivity of advection to velocity errors.
Against purely data-driven baselines on the same synthetic setting, UNet2 has test MAE 0.1314 °C, UNet3 has 0.0473 °C, and DDUNet4 has 0.0548 °C (Pelzer et al., 8 Jul 2025). The reported point is not only that LGCNN attains similar or better accuracy, but that it does so with far less training data: roughly 1 training datapoint in the 3dp setup, versus 73 training datapoints for the 101dp baselines. On real permeability fields from Munich, the isolated Step 3 validation MAE is 5, the full pipeline validation MAE 6, and the scaling full pipeline MAE 7.
Adjacent local-global CNN formulations provide corroborating evidence that explicit global augmentation can improve performance. In image enhancement, the proposed local-global network reports average scores over 283 test images of TMQI 0.8556, entropy 6.4126, NIQE 4.7207, and BRISQUE 45.7993, and the paper states that the global encoder suppresses distortions seen without it (Kinoshita et al., 2019). In local feature extraction, adding the Global Enhancement Module changes HPatches metrics from Rep 0.6099, MS 0.4173, MMA 69.21%, MHA 70.93%, Params 0.3292M to Rep 0.6178, MS 0.4185, MMA 70.90%, MHA 71.30%, Params 0.3625M, while maintaining GLFeat FPS = 153.39 (Miao et al., 2022). In classification, the GLP model improves Caltech101 Top-1 accuracy from 86.41 to 91.24 for ResNet18 to GA-Resnet and from 90.50 to 92.40 for InceptionV3 to GA-Inception (Rezvani et al., 2023).
These results should not be conflated as measurements of a single standardized LGCNN benchmark. Rather, they show that explicit separation of local and global processing is effective across multiple tasks, even though the global mechanism varies substantially.
7. Limitations, assumptions, and conceptual significance
The explicit LGCNN paper states several limitations. The current approach is limited to 2D and steady-state datasets, relies on offline labeled simulations for training, and uses a simplified transport surrogate rather than a full PDE solver (Pelzer et al., 8 Jul 2025). It also assumes that heat transport is approximated as a one-way coupled process, that the flow field is steady during training and inference, and that streamlines sufficiently capture the dominant nonlocal behavior.
These caveats are important because they delimit what LGCNN means in this context. The architecture is not a universal substitute for physics simulation; it is a modular surrogate tailored to problems in which some sub-processes are locally learnable while others are better represented explicitly. This suggests a broader methodological lesson: in scientific ML, “global” need not be another neural attention layer. It can be an external operator, solver, or prior that exposes the nonlocal structure to the network.
A second source of ambiguity concerns naming. Some papers are clearly LGCNN-like in design but do not use the acronym, while others use nearby terms such as “global-local model,” “localized CNN,” or “Global/Local Processing” (Lu et al., 2024, Uselis et al., 2020, Rezvani et al., 2023). In nuclear mass prediction, for example, the CNN-WS4 model combines a local CNN over a 8 neighborhood with the global WS4 mass model by learning the residual
9
which is described as a global-local fusion model rather than LGCNN proper (Lu et al., 2024). This suggests that LGCNN is better treated as a family resemblance concept than as a fully standardized taxonomy.
Taken together, the literature supports a precise characterization. LGCNN refers most strictly to architectures that preserve the local inductive bias of convolution while introducing an explicit global mechanism—solver-based, encoder-based, attention-based, or position-dependent—to handle dependencies that local receptive fields alone represent poorly. In that sense, the groundwater formulation is the clearest named exemplar (Pelzer et al., 8 Jul 2025), while the surrounding literature demonstrates the breadth of the local-global design principle across applied machine learning.
