Single-well models (also referred to as local or single-station models in the groundwater literature) are trained individually for each monitoring well, based on the assumption that each time series originates from its own data-generating process . These models enable a detailed representation of local hydrogeological characteristics and dynamic changes , offering high interpretability due to their sensitivity to site-specific input features and time windows. However, their applicability is primarily limited by a tendency toward overfitting and a lack of generalizability to other locations, as models must be trained separately for each well. Consequently, local models cannot exploit spatial variability or regional dynamics present in different time series across the monitoring network .

Global models (also referred to as regional or multi-well/multi-site models) are trained on combined data from multiple monitoring wells and can generate predictions for all locations included in the training set . This approach enables efficient use of large datasets and facilitates information sharing across the entire network . Owing to their architecture, global LSTM models can identify and generalize co-occurring patterns across time series . A key motivation is spatial transfer, i.e., applying a model trained on many wells to entirely withheld wells (spatial out-of-sample sites) when suitable static descriptors are available; in streamflow hydrology, this is closely related to the PUB paradigm (Prediction in Ungauged Basins). Streamflow studies have shown that global models can improve generalization across basins, including transferability to ungauged or data-sparse basins and unseen periods . For groundwater head prediction, however, systematic large-sample evidence for spatial out-of-sample generalization remains limited, and recent work suggests that spatial transfer can be challenging , partly because static attributes may act primarily as identifiers rather than enabling transferable representations . Additional benefits include the ability to model long-memory patterns, robustness to data gaps, and potentially higher computational efficiency compared to training many local models individually . In surface-water benchmarking studies, global models have frequently outperformed conventional, locally calibrated hydrological models .

However, the often-assumed superiority of global models is increasingly questioned. In a large-scale evaluation, found that a Google-developed global streamflow model (trained on 5680 basins) underperformed locally trained single-basin models in 46 % of 609 catchments and substantially underestimated high flows (95th–99th percentiles) by an average of 45 %. Consistently, their meta-analysis of 123 studies showed that single-basin models frequently attain high skill (NSE ≥ 0.75 in >92 % of cases), underscoring that local models can be highly competitive.

Although this evidence comes from the surface-water domain, similar limitations have been reported for globally trained models in groundwater head prediction. In the presence of highly heterogeneous groundwater head time series, performance can decline . Global models tend to focus on dominant shared patterns, potentially at the expense of local variability. Learning nonlinear relationships between inputs and targets can become challenging when fundamentally different dynamical behaviors are combined during training . Furthermore, studies have shown that static features such as geology, climate, or land use often fail to create proper entity awareness and instead act merely as identifiers , which limits spatial generalization. Deficits have also been observed for extreme groundwater conditions, for example, due to saturation effects in the LSTM architecture or underestimation of extremes . More generally, sequence models such as LSTMs rely on bounded gating and activation functions, which can limit extrapolation beyond the training range and contribute to biased predictions under unprecedented extremes . Finally, the black-box nature of deep neural networks remains a key challenge for decision support in groundwater management, especially for global models, as they capture complex, cross-site patterns that reduce the transparency of local relationships .

Partitioned models (also referred to as clustering-based or subgroup-specific models) are essentially global models trained on subgroups of monitoring wells that share similar temporal dynamics or static attributes. These models operate on more homogeneous subgroups of time series, which are typically formed through data-driven methods (e.g., clustering algorithms) or domain-specific groupings. The objective is to homogenize training data and to specifically align modeling capacity with similar time series types . Clustering is usually based on (i) dynamic time series features such as trend, seasonality, autocorrelation, or entropy ; (ii) spectral characteristics to separate typical frequency patterns ; (iii) shape-based similarity metrics such as Grey Relational Analysis (GRA) ; or (iv) static site attributes such as climate, geology, or topography . In the surface-water domain, analogous groupings can also be based on static catchment attributes . Subsequently, a dedicated model is trained for each group. Several studies have shown that partitioned models are often more robust to heterogeneity than fully global approaches, particularly when time series exhibit strongly divergent dynamics . By focusing on homogeneous subgroups, partitioned models can enhance both predictive performance and interpretability .

In light of these developments, this study aims to systematically compare the predictive performance of global and local deep learning (DL) models for groundwater level forecasting. The central question is whether the advantages of globally trained models, whose superior performance has been widely demonstrated in hydrological streamflow modeling, can be transferred to hydrogeological applications, particularly under the specific conditions of diverse system dynamics; ranging from highly dynamic behavior in karstic aquifers to inertial responses in low-permeability porous aquifers; as well as heterogeneous site conditions and limited data availability.

In contrast to previous studies, the analysis is based on an extensive, Germany-wide groundwater level benchmark dataset comprising nearly 3000 monitoring wells , spanning over three decades. The associated spatial and dynamic diversity enables a differentiated assessment of the generalizability of data-driven models across different geological and climatic settings.

The core research questions are:

Overall Model Performance: Are globally trained LSTM models generally superior to local (single-well) models in terms of overall predictive accuracy across a large and heterogeneous set of monitoring wells?

To address these questions, we conducted a comprehensive experimental comparison. The experiments involve global LSTM models, trained either on the full dataset or on differently partitioned subsets, and locally trained CNN single-well models. All models are evaluated on the same standardized data basis, using test designs that systematically vary the size and dynamic composition of the training dataset, as well as extrapolative settings, including extreme groundwater levels outside the training range (below the well-specific 1st percentile or above the 99th percentile of the training distribution) and spatial out-of-sample prediction at monitoring wells withheld from training.

The analysis is based on the GEMS-GER dataset , which provides standardized groundwater-level observations and associated predictor variables for Germany. The dataset contains weekly time series from 3207 monitoring wells for 1991–2022, covering all major hydrogeological regions and a wide range of aquifer types and system dynamics. For this study, we used a filtered subset of 2951 wells, excluding sites that achieved an NSE ≤0 across all three benchmark baseline models provided with the GEMS-GER benchmark workflow (single-well CNN and global LSTM). Details on data sources, preprocessing, and quality control are provided in the dataset description. The full dataset is openly available via Zenodo (DOI: 10.5281/zenodo.15530171, ). The spatial distribution of the monitoring wells is shown in Appendix A (Fig. ), and representative time series are shown in Appendix B (Fig. ).

Each groundwater time series is complemented by dynamic input variables representing meteorological and hydrological conditions, including precipitation, temperature, relative humidity, evapotranspiration, soil moisture, soil temperature, snowmelt, snow water equivalent, and surface as well as subsurface runoff. These variables are taken from the GEMS-GER dataset and were used here exactly as defined therein . All dynamic inputs are provided at weekly resolution; daily fields were aggregated to weekly values using variable-specific operators (weekly mean or weekly sum depending on the variable; see Table 1 in the GEMS-GER dataset paper). We did not compute additional multi-window indices or running aggregations (e.g., SPI or rolling net-precipitation sums) beyond the weekly aggregation already defined in GEMS-GER. HYRAS-based variables were selected whenever available (higher spatial resolution), and ERA5-Land was only used to complement variables not provided by HYRAS.

In addition to dynamic inputs, each monitoring well is characterized by a set of more than 50 static attributes, including hydrogeological, topographic, soil, and land-use properties. Static attributes are time-invariant site descriptors and do not include statistics derived from the groundwater-level time series (e.g., mean head or standard deviation). From the full set of static features provided in the dataset , variables related to well depth, screen characteristics, pumping, and pressure state were excluded, as these were sparsely available for the majority of monitoring wells. All categorical static features were label-encoded for use in the machine-learning models.

We implemented and compared two main types of modeling strategies: local (single-well) models and global models. The latter were trained either on the full dataset or on differently partitioned subsets (referred to as partitioned models). Throughout this study, we refer to local (single-well) models as S, global models as G, and partitioned variants as S-Px and G-Px, where x indicates the partition stage. To indicate the partitioning strategy, the subscript COR denotes correlation-based removal (increasing dynamic similarity), and the subscript RND denotes random removal.

(i) Local (Single-Well) Models (S-P0): Independent CNN models were trained for each monitoring well using only local dynamic input variables. These models serve as a site-specific baseline without transferring cross-site information. All single-well models were trained for each of the 2951 wells (Stage P0).

The partitioning procedure is defined as follows:

Stages P1–P5_COR: Starting from P0, an additional 500 wells were successively removed in each stage based on their dynamic dissimilarity to other wells. To quantify this, we computed the pairwise absolute Pearson correlations between the standardized groundwater level time series. Each well’s dynamic representativeness was then defined as the mean absolute correlation with all others. Wells with the lowest representativeness were considered least typical in terms of dynamics and removed first, resulting in subsets with increasing internal similarity. We deliberately did not impose a spatial constraint on the similarity criterion, as dynamically similar groundwater responses are not necessarily local and preserving such non-local analogs is part of the motivation for global learning. While spatio-dynamic clustering is a plausible alternative, it introduces additional design choices (e.g., spatial weighting and cluster definition) and would make the controlled, stage-wise comparability of the progressive reduction (P0–P5) less straightforward.

All models in this study follow the benchmark architectures introduced in , using a standard sequence-to-value forecasting setup. Input sequences of 52 time steps (i.e., weeks) were used to predict the groundwater level at the following time step. Models were trained and validated on the periods 1991–2007 and 2008–2012, respectively, and evaluated on the final 10 years (2013–2022). All metrics were computed from the median prediction of an ensemble of ten independently initialized models.

In this study, single-well models are implemented using a 1D convolutional neural network (CNN), whereas global models are implemented using a Long Short-Term Memory network (LSTM). Architecture selection was guided by preliminary baseline experiments and computational practicality, as the study aims to isolate the effects of training strategy rather than to benchmark architectures. For single-well training across thousands of wells, the CNN yielded comparable skill while being markedly faster and more stable (LSTM runs were more prone to optimization instabilities), whereas the LSTM provided a strong and widely used baseline for global multi-well sequence learning. To enable a controlled comparison across training strategies and partitioning stages, we adopted the benchmark architectures and hyperparameter settings of and kept them fixed throughout; robustness to stochasticity is addressed via an ensemble of ten initializations (median performance). All dynamic inputs and the target groundwater head series were standardized per well (z-score) using pre-test statistics (1991–2012) and back-transformed accordingly. In the spatial out-of-sample experiments, wells were withheld from weight training, but their scaling parameters were derived from their own pre-test head observations.

The single-well models are based on a 1D convolutional neural network (CNN) architecture. Each model consists of a convolutional layer with 256 filters and kernel size 3, followed by max pooling, flattening, a dense layer with 32 units, and a final output layer. The models were trained using the Adam optimizer (learning rate 0.001), early stopping (patience 5), a batch size of 16, and a maximum of 30 epochs. Only dynamic input features were used.

The global models are based on a Long Short-Term Memory (LSTM) architecture. The dynamic input branch consists of a single LSTM layer with 128 units, followed by a dropout layer with a dropout rate of 0.3. Models were trained for up to 20 epochs using a batch size of 512, early stopping (patience: 5), and a learning rate scheduler targeting a value of 0.001. Static features were incorporated using a second model branch that processes static inputs via a dense layer with 128 units. The outputs of both branches are concatenated and passed through a dense layer with 256 units before the final output layer. Categorical static features were label-encoded. For further architectural and implementation details, we refer to .

All global models (G and G-Px) were retrained independently using the same LSTM architecture and hyperparameters, ensuring architectural consistency across all partitioning stages. In contrast, the single-well models (S) were trained once per well on the P0 dataset and remained unchanged; for each partition, only the models corresponding to the retained wells were considered in the evaluation.

The experimental design addresses the four research questions outlined in Sect. , each examining a distinct aspect of model performance and generalization. To systematically evaluate these aspects, we conducted four targeted experiments focusing on:

(i) Overall Performance Comparison: We compared the predictive accuracy of global, local, and partitioned models across all monitoring wells (P0). This experiment serves as a baseline to assess overall model performance and consistency across dynamic groundwater regimes.

To assess whether globally trained LSTM models outperform local single-well (S-P0) models, we compare their predictive performance across 2951 monitoring wells. Both models were trained on the full dataset (P0).

Figure a compares global (G-P0) and single-well (S-P0) performance. Overall, differences are moderate: the median NSE is slightly higher for S-P0 (0.49 vs. 0.47; Table ), and S-P0 attains more high-performing wells (upper tail). G-P0 shows a slightly narrower interquartile range, but also more very low NSE values, indicating an elevated risk of underperformance at some locations.

The pairwise comparison in Table  confirms these trade-offs: G-P0 outperforms S-P0 at 45.4 % of wells, underperforms at 48.9 %, and performs equally (within ±0.01 NSE) at 5.7 %. Thus, both model types perform broadly comparably, reflecting the balance between generalization and local adaptation.

The cumulative NSE curves (Fig. a, lower) show that performance differences vary across the NSE spectrum: S-P0 is slightly better at the very low end (NSE < 0.05), G-P0 performs better from 0.05 to 0.325, both are similar between 0.325 and 0.425, and S-P0 is more favorable in the mid-to-high NSE range.

To test whether these local differences are spatially structured, Fig.  maps per-well ΔNSE=NSEG-NSES and summarizes local spatial association using a LISA analysis. Although ΔNSE is close to zero on average (P0: mean =-0.012, median =-0.006, share of wells with Δ>0 is 48.7 %), it exhibits significant positive spatial autocorrelation (Global Moran's I=0.322, p=0.001). The LISA results indicate localized clusters of consistently positive or negative ΔNSE as well as spatial contrasts (19.7 % significant at α=0.05), with the most pronounced clustering in hydrogeological region (3) (Upper Rhine Graben including the Mainz Basin and the North Hessian Tertiary), where dense monitoring facilitates the detection of significant local patterns and the patchwork of river-influenced and more regionally coherent dynamics likely contributes to spatially organized model advantages. Overall, this suggests that performance differences are spatially organized in parts of the domain, while remaining non-significant for the majority of wells.

To assess how training data characteristics affect global model performance, we analyze three complementary experiments: (i) sensitivity to training-record length by progressively shortening the available training period in annual steps while keeping the validation and test periods fixed (2008–2012 and 2013–2022), (ii) increasing dynamic similarity through correlation-based well removal, and (iii) reducing training data volume (while maintaining diverse dynamics) through random subsampling.

To assess model performance under extrapolative conditions, we evaluated predictions for groundwater levels (GWLs) beyond the typical range observed during training. For each well, low extremes were defined as values in the test period below the 1st percentile of its training distribution, and high extremes as values above the 99th percentile. This site-specific percentile approach ensures that extremes are identified relative to each well’s training history, while avoiding dependence on absolute thresholds.

Figure  summarizes RMSE distributions for all extrapolated values (top), low extremes (middle), and high extremes (bottom), using both correlation-based and random partitioning. Across all stages, global models do not show improved predictive skill over single-well models. For low extremes, errors are slightly higher for global models at every stage, suggesting that dynamics associated with exceptionally low GWLs are underrepresented in the training sets. For high extremes, both model types perform similarly, with neither showing a consistent advantage.

The stability of error distributions across increasing training set homogeneity or size indicates that, in groundwater systems, larger or more homogeneous datasets do not automatically enhance the prediction of extremes. A plausible explanation is that extreme events often depend on site-specific factors such as fine-scale geology, localized abstraction, or land use, which are not fully captured by the available static site descriptors. Without sufficiently informative descriptors, the transfer of extreme-event knowledge between sites is limited, and events not directly inferable from the dynamic meteorological inputs remain difficult to predict. This reflects a general constraint of current large-scale groundwater datasets rather than a shortcoming of the modeling approach itself.

To evaluate the spatial transferability of global models, we assessed their performance on monitoring wells deliberately excluded from model calibration. This simulates predictions at sites without prior training data, while observations remain available for evaluation. Out-of-sample (OOS) subsets were defined using the partitioning strategies introduced in Sect. , i.e., correlation-based removal of dynamically dissimilar wells and random exclusion. Model performance at these OOS sites was compared to that of single-well models trained individually on each target well (in-sample reference). Figure  summarizes the resulting differences in predictive skill.

At stage P_x, the global model is trained on the remaining wells of that stage and evaluated on the wells excluded up to that stage (cumulative OOS target set). Hence, the OOS test set varies across stages (increasing from 500 to 2500 wells), and cross-stage comparisons should be interpreted as transfer to different target populations rather than a like-for-like evaluation on a fixed test set.

Our findings provide mixed support for earlier results from deep learning applications in hydrology and hydrogeology. In line with studies in streamflow modeling , we find that global models can achieve predictive skill comparable to or exceeding that of local models when trained on large dynamically homogeneous datasets. This confirms the general advantage of cross-site learning in environments where system dynamics are similar, as also observed in other partitioning approaches . More generally, multi-site training can be attractive because it consolidates model development into a single network-level model (instead of thousands of per-well calibrations) and, in principle, enlarges the training envelope by exposing the model to a broader range of hydro-climatic situations and response regimes. This is often discussed as a pathway to improved robustness and information sharing, particularly for sites with limited local data.

When the available training history is progressively shortened while validation and test remain fixed, single-well model performance deteriorates more strongly and becomes more variable, whereas the global model remains comparatively stable across record lengths. This pattern is consistent with the notion that local deep-learning models require sufficiently long site records to reach their full potential, while global training can partially compensate limited local information via cross-site learning . In other words, even if global models do not provide a systematic advantage under heterogeneous conditions for long records, their relative robustness under shorter records supports the relevance of multi-site learning for applications where monitoring histories are limited.

However, while global models have often shown clear advantages for streamflow applications , though not without recent dissenting findings , our results for groundwater level prediction reveal no overall advantage under heterogeneous conditions. This echoes the broader debate that “global-model superiority” is not guaranteed: even in hydrology, global models can underperform local ones when local dynamics are highly site-specific and when local data quality or representativeness exceeds what the shared feature space can explain.

This difference is likely related to the broader diversity and high small-scale variability of groundwater system dynamics, ranging from highly responsive karst aquifers to inertial systems in low-permeability sediments. Unlike many surface water catchments, groundwater dynamics can differ markedly even over short distances. Nearby wells may share similar static feature values (e.g., geology, land use) yet exhibit distinct responses due to fine-scale geological differences, flow paths, or localized abstraction.

Under such conditions, a global model may still learn a broad range of dynamics present in the training set but (because the available static site descriptors are not sufficiently informative to uniquely identify the controlling local conditions) cannot reliably assign the correct dynamic regime to a specific well. This is the case even though our benchmark includes a comparatively rich set of >50 time-invariant static attributes (hydrogeology, topography, soils, land use) and intentionally excludes groundwater time-series statistics. As a result, the model tends to average across similar but not identical behaviors, which reduces site-specific accuracy under heterogeneous conditions. This highlights a key trade-off between global and local strategies: global models emphasize breadth and generalization, whereas single-well models emphasize local precision and are less prone to “smoothing” unique site behavior into a dominant average regime. In settings with strong local controls that are not fully observable from static descriptors, local models can remain highly competitive – especially when long, high-quality site records are available. This mechanism is consistent with the strong empirical link between time-series representativeness and model skill observed in our experiments: wells whose dynamics are well represented by the remaining training pool (high mean absolute correlation) are systematically easier to predict, whereas atypical wells show substantially higher error variance. Targeted correlation-based filtering shifts the training data towards this representative regime and thereby reduces conflicting learning signals. Consistent with the concerns raised by , this reflects a broader limitation of large-scale groundwater applications, where even rich static descriptors may not fully capture fine-scale controls such as local flow paths or abstraction. While dynamic similarity may correlate with some static attributes, our filtering is purely time-series based and not spatially constrained; therefore, increasing dynamic similarity does not necessarily imply a strong homogenization of static features.

Finally, our partitioning experiments confirm the robustness benefits reported in other hydrological contexts : grouping wells with similar dynamics before training significantly improved the performance of global models, even with fewer training wells. Notably, the ΔNSE diagnostics further indicate that these gains are not merely a consequence of evaluating on an increasingly “easier” subset: when assessed on a fixed well set (the final P5 wells), correlation-based reduction yields predominantly positive performance differences relative to earlier stages and converges towards zero in later stages, whereas random reduction remains broadly centered around zero with substantial site-dependent gains and losses. Together, these results reinforce the broader conclusion from the literature that data homogeneity, whether achieved via targeted filtering or clustering, can be more important for generalization skill than sheer dataset size. In groundwater modeling, dynamic similarity therefore often outweighs data quantity as a determinant of global model skill. In this sense, partitioned (or clustered) global models can be viewed as a pragmatic middle ground: they retain some benefits of information sharing within dynamically coherent subgroups while reducing the risk that strongly dissimilar wells impose conflicting learning signals that hamper site-specific accuracy.

Not necessarily. Despite being trained on a large and diverse dataset, globally trained models did not show a notable overall advantage over locally optimized single-well models. Local models tend to achieve slightly higher median accuracy and perform better at individual sites, while global models produce more predictions clustered around the central range of performance, suggesting a more stable but less specialized behavior. This pattern is consistent with limited entity awareness under heterogeneous conditions, where even a rich set of time-invariant static attributes (>50 hydrogeological, topographic, soil, and land-use descriptors; no groundwater time-series statistics) may be insufficient to uniquely identify the controlling local processes at each well. Conceptually, this reflects the core trade-off: global models emphasize breadth and potential robustness via information sharing, whereas single-well models emphasize local precision and can better capture unique site-specific behavior when long, high-quality local records are available. These findings align with earlier work in groundwater modeling but contrast with the consistent superiority reported for streamflow, likely reflecting the greater diversity and small-scale variability of groundwater system dynamics. However, when training is restricted to dynamically homogeneous subsets, global models can match or even exceed single-well performance at many sites, highlighting that cross-site learning becomes beneficial once heterogeneity is reduced.

Training data quality in terms of dynamic similarity is more important than quantity. When the training set is filtered to include only wells with similar temporal dynamics, global model accuracy improves markedly, and performance changes become more consistent across sites. In contrast, random removal of wells (reducing size without regard to similarity) does not yield comparable gains, even when roughly 85 % of wells are removed; any changes remain modest and site-dependent. Importantly, evaluating all stages on a fixed test set (the final P5 wells) confirms that the improvements under correlation-based filtering are not merely a consequence of progressively excluding difficult sites, but reflect a genuine benefit of training on a more dynamically consistent data basis. Thus, for in-sample prediction, structure (dynamic similarity/representativeness) matters more than sheer sample size; simply enlarging the dataset without improving consistency provides little benefit. For in-sample predictions, dynamic similarity is clearly the dominant factor; for out-of-sample predictions, a broader and more diverse training base may, in principle, offer robustness benefits, although this was not a dominant effect in our experiments. Consistent with this, global models were also less sensitive to shorter training histories than single-well models under our benchmark setup.

No. In our experiments, neither global nor local models could reliably predict groundwater levels outside the training range. Across all partitioning stages and extrapolation regimes, global models showed slightly higher errors and a tendency to overestimate low values while underestimating high ones, consistent with a structural averaging effect. Neither increasing the amount of training data nor improving dynamic similarity mitigated this issue. A likely reason is that the available static site descriptors (e.g., geology, land use, geomorphology) are not sufficiently informative to provide strong entity awareness and to capture the local controls governing extremes, thereby limiting the transfer of extreme-event knowledge between sites. These descriptors represent the best practical dataset currently available for large-scale groundwater modeling, so this limitation reflects a general constraint of current data availability rather than a shortcoming of the modeling approach itself.

Under correlation-based exclusion (i.e., deliberately held-out wells with low dynamic similarity), the global model shows limited spatial generalization and performance drops sharply compared to single-well models, reflecting the difficulty of transferring learned dynamics to dissimilar sites. Across successive stages, the OOS target set expands from the most atypical wells to include progressively less dissimilar sites, which leads to a modest rightward shift in global performance; however, the gap to single-well models remains substantial. In contrast, random exclusion yields broadly similar performance across stages, indicating that generalization is feasible when target wells share representative temporal patterns with the training data, consistent with previous findings that groundwater dynamics are less transferable than streamflow. Importantly, our OOS setup evaluates transfer to wells with available historical observations (needed for well-specific scaling), i.e., a realistic “new well with records but not used for training” setting rather than a fully no-head-data scenario.

In summary, the choice between global and local modeling depends strongly on the intended application. For in-sample prediction at known sites, training on dynamically similar wells (e.g., by partitioning or filtering the dataset based on dynamic similarity) can yield accurate results even with relatively few data, as the remaining wells benefit from increased dynamic similarity and the reduced influence of poorly representative sites. For spatial generalization, a broad and diverse training base may improve robustness, although this effect was modest in our experiments and can come at the cost of predictive precision. From an operational perspective, global models can still be attractive because they provide a single maintainable model for an entire monitoring network and enable weight transfer across sites; however, where the goal is maximal site-specific accuracy (especially for atypical wells), single-well models remain a strong and often preferable option. For groundwater systems, characterized by slow and often indirect responses, sparse measurements, high small-scale variability, and limited entity awareness due to coarse static descriptors, model transferability appears inherently more limited than in surface water applications. This highlights that single-well models remain a strong option, especially when site-specific accuracy is required and local dynamics need to be captured in detail.