Modelling Future Evolution of the Antarctic Ice Sheet With Explainable Machine Learning Authors: Paula Maddigan, Bach Hoai Nguyen, Peter Yu Feng Siew, Bing Xue, Nicholas Golledge, Mengjie Zhang Published: 2026 ABSTRACT The accelerating effects of climate change, particularly rising sea levels resulting from melting land ice, require accurate predictions of ice sheet dynamics. Traditional numerical models, such as the Parallel Ice Sheet Model (PISM), are widely used; however, they rely on simplifying assumptions and are computationally intensive, limiting their ability to make long‐term predictions. This study addresses these limitations by employing machine learning techniques to emulate ice sheet dynamics, offering a faster, more efficient alternative to traditional numerical models. Using data generated by PISM for the Antarctic ice sheet spanning 2015 to 2100, we develop proof‐of‐concept machine learning models that predict key ice properties—ice mask, ice thickness, and ice velocity—based on climate inputs. Our models tackle both spatial and temporal challenges by incorporating the ice properties of neighbouring cells and past temporal values, thereby improving predictive accuracy. Additionally, explainable AI (XAI) techniques are applied to enhance trust and explainability. Our best‐performing model, a Random Forest with an expanding window, achieves an average RMSE of 3.36 for thickness and 65.6 for velocity and misclassifies on average only 3 grid points per year for the ice mask out of 2257, demonstrating good accuracy comparable to that of the original simulator at a fraction of the computational cost. This approach not only improves the computational efficiency of scenario modelling but also contributes to the broader field of climate science by enabling more frequent and extensive analyses of climate change impacts on Antarctic ice sheets. Introduction Climate change is leading to a concerning rise in sea levels (Golledge 2020; Levy et al. 2023; Hanna et al. 2024), with the melting of land ice being a primary contributor (Edwards et al. 2021). Since the 1990s, Greenland and Antarctic ice loss have contributed about 21 mm to global mean sea level rise (Otosaka et al. 2022). In the future, Antarctic ice loss is expected to be the largest contributor, reaching up to 34 cm rise by 2100 and several metres by 2300 under the high‐emission Shared Socioeconomic Pathway 5–8.5 (Fox‐Kemper et al. 2021). It is essential to gain insight into the future impacts of climate variability and global warming on ice sheets to develop effective strategies to mitigate the impacts of ongoing ice loss. Policymakers need accurate and efficient projections of ice loss to develop adaptation plans to minimise the social and economic impacts of sea‐level rise, such as relocating assets and communities and building flood‐resistant infrastructure in low‐lying coastal areas (Magnan et al. 2022). Considerable research has focused on developing numerical models to simulate ice sheets (Goelzer et al. 2020; Seroussi et al. 2024). However, most require assumptions to simplify complex physical processes. Furthermore, the high computational cost makes it challenging for existing numerical models to produce large numbers of long‐term (multi‐millennial) predictions necessary for robust uncertainty quantification. To address this issue, we develop a proof‐of‐concept 2D machine learning (ML) emulator to simulate the spatial evolution of ice sheet thickness and flow. ML identifies complex patterns, offering an alternative to conventional numerical models that rely on intricate physical computations. By training models on historical data using advanced algorithms, ML approaches can provide timely predictions without compromising accuracy. This capability is particularly valuable for scenario‐based ensemble modelling, offering a faster alternative to traditional methods. Previous studies have explored the development of statistical emulators to better understand the future contributions of Antarctic ice loss to global sea‐level rise (Berdahl et al. 2020; Lowry et al. 2021; Edwards et al. 2021; DeConto et al. 2021). However, an explainable AI (XAI) emulator that captures the spatial–temporal variability of both the thickness and the dynamic changes of the Antarctic ice sheet has not yet been achieved. The primary goal of this study is to emulate a low‐resolution version of a previously published simulation (Golledge et al. 2015) that used the Parallel Ice Sheet Model (PISM) (Bueler and Brown 2009). The aim is to achieve greater efficiency than the numerical model while making fewer assumptions. To accomplish this, we develop ML models using PISM‐generated ice sheet data for Antarctica spanning 86 years (2015 to 2100). The computational domain is composed of equidimensional (x, y) cells, each with associated coordinates. For each year in the dataset, every cell has three climate variables: air temperature, ocean temperature, and precipitation. We aim to predict three outputs: ice mask, ice thickness, and ice velocity. Further details on the three outputs are presented in Section 3.1 and summarised in Table 1. There are two main challenges in deploying ML‐based ice sheet models: TABLE 1Descriptions of the dataset.MetadataDescriptionYearSimulation calendar year (2015 to 2100)XX coordinate of the grid pointYY coordinate of the grid point Climate variables Description PrecipitationMeasurement of rain, hail, sleet, snow (mm/year)Air temperatureMeasurement of air temperature (Kelvins)Ocean temperatureMeasurement of ocean temperature (Kelvins) Target variables Description Ice mask2 (grounded ice), 3 (floating ice), or 4 (open ocean)Ice thicknessThickness of ice (metres)Ice velocitySpeed of ice movement (metres/year)© 2026 WILEY‐VCH GmbH Descriptions of the dataset. •Spatial challenge: The three ice properties (ice mask, thickness, and velocity) of each cell also depend on the climate variables and the ice properties of the neighbouring cells. Thus, to ensure good predictive performance, we will need to consider neighbouring cells in our models.•Temporal challenge: The current ice properties are dependent on the ice properties from previous years. More specifically, predicting ice properties is inherently a time‐series problem, requiring the incorporation of past values into our models to achieve more accurate predictions. •Spatial challenge: The three ice properties (ice mask, thickness, and velocity) of each cell also depend on the climate variables and the ice properties of the neighbouring cells. Thus, to ensure good predictive performance, we will need to consider neighbouring cells in our models. Spatial challenge: The three ice properties (ice mask, thickness, and velocity) of each cell also depend on the climate variables and the ice properties of the neighbouring cells. Thus, to ensure good predictive performance, we will need to consider neighbouring cells in our models. •Temporal challenge: The current ice properties are dependent on the ice properties from previous years. More specifically, predicting ice properties is inherently a time‐series problem, requiring the incorporation of past values into our models to achieve more accurate predictions. Temporal challenge: The current ice properties are dependent on the ice properties from previous years. More specifically, predicting ice properties is inherently a time‐series problem, requiring the incorporation of past values into our models to achieve more accurate predictions. We conducted a comprehensive empirical analysis to assess the validity and accuracy of our models. To establish trust in our approach, we applied XAI techniques that provide both local and global explanations for ice sheet dynamics predictions. By combining ML methods with data from physical climate models, our models achieve high accuracy and efficiency while remaining interpretable. By formulating the ice sheet problem as a ML task, we address key challenges related to spatial and temporal dependencies through feature engineering. Our contributions are as follows: •Problem formulation: We reframe ice sheet modelling as a ML task, structuring the prediction problem to capture both spatial and temporal dependencies in climate and ice sheet dynamics.•Feature engineering for spatio‐temporal learning: We analyse PISM‐generated data and construct new features that encode long‐term memory effects and climate interactions to enhance learning performance.•Learning scenarios: We propose two learning scenarios: (1) one‐step ahead forecasting where we update the model annually with historical data and (2) multi‐step forecasting where we use past predictions to update the model to make future predictions. These scenarios explore different ways to incorporate past climate and ice data, offering insights into how historical dependencies affect prediction accuracy.•Model explainability: We conduct a detailed explainability analysis by using Random Forest feature importance and SHAP (SHapley Additive exPlanations) (Lundberg and Lee 2017) toolset. This analysis highlights the contributions of past climate data, air temperature, ocean temperature, and precipitation to ice dynamics predictions. •Problem formulation: We reframe ice sheet modelling as a ML task, structuring the prediction problem to capture both spatial and temporal dependencies in climate and ice sheet dynamics. Problem formulation: We reframe ice sheet modelling as a ML task, structuring the prediction problem to capture both spatial and temporal dependencies in climate and ice sheet dynamics. •Feature engineering for spatio‐temporal learning: We analyse PISM‐generated data and construct new features that encode long‐term memory effects and climate interactions to enhance learning performance. Feature engineering for spatio‐temporal learning: We analyse PISM‐generated data and construct new features that encode long‐term memory effects and climate interactions to enhance learning performance. •Learning scenarios: We propose two learning scenarios: (1) one‐step ahead forecasting where we update the model annually with historical data and (2) multi‐step forecasting where we use past predictions to update the model to make future predictions. These scenarios explore different ways to incorporate past climate and ice data, offering insights into how historical dependencies affect prediction accuracy. Learning scenarios: We propose two learning scenarios: (1) one‐step ahead forecasting where we update the model annually with historical data and (2) multi‐step forecasting where we use past predictions to update the model to make future predictions. These scenarios explore different ways to incorporate past climate and ice data, offering insights into how historical dependencies affect prediction accuracy. •Model explainability: We conduct a detailed explainability analysis by using Random Forest feature importance and SHAP (SHapley Additive exPlanations) (Lundberg and Lee 2017) toolset. This analysis highlights the contributions of past climate data, air temperature, ocean temperature, and precipitation to ice dynamics predictions. Model explainability: We conduct a detailed explainability analysis by using Random Forest feature importance and SHAP (SHapley Additive exPlanations) (Lundberg and Lee 2017) toolset. This analysis highlights the contributions of past climate data, air temperature, ocean temperature, and precipitation to ice dynamics predictions. Related Work Existing studies present a range of approaches to simulating the complex dynamics of the Antarctic ice sheet, from empirical models to more complex physical models. Empirical statistical models (Rahmstorf 2007; Vermeer and Rahmstorf 2009; Levermann et al. 2013; Schaeffer et al. 2012) use historical data and observed trends with linear relationships to predict future changes. While relatively simple and easy to implement, their heavy reliance on historical data means that these models may not adequately capture complex non‐linear interactions and the long‐term trends associated with future climate change. Physical models such as the PISM (Bueler and Brown 2009; Winkelmann et al. 2011), the Ice Sheet System Model (ISSM) (Larour et al. 2012), and the Community Ice Sheet Model (CISM) (Lipscomb et al. 2019) 1 are extensively used for understanding the response of the ice sheet to external factors. They encapsulate the physics of ice dynamics and their interactions with environmental indicators, enabling long‐term predictions and accommodating complex non‐linear relationships. Their design allows for scenario modelling, an important tool for analysing the impact of climate change in Antarctica. Recent advancements in artificial intelligence (AI) have led to ML solutions that efficiently process large, high‐dimensional datasets and capture complex relationships. A proof‐of‐concept study (Hu et al. 2021) developed a multi‐layer perceptron (MLP) model for surface melt over the Larsen Ice Shelf, outperforming simulations from the regional climate model RACMO2. Deep learning methods such as Convolutional Neural Networks (CNNs) and Physics‐Informed Neural Networks (PINNs) have become popular in climate science, capturing spatio‐temporal patterns and building physical constraints into the learning process. For example, CNNs have been used to emulate complex climate fields, while PINNs have been developed into general purpose libraries, such as PINNICLE v1.0, to solve PDE‐constrained problems (Cheng et al. 2025). Similar ideas have also been explored in related areas, such as accelerating subglacial hydrology with neural networks (Verjans and Robel 2024) and blending ML with dynamical climate forecasts (Slater et al. 2023; Liu et al. 2025). By comparison, ice sheet emulation has received much less attention, partly due to the unique challenges posed by long timescales and strong memory effects, in which past changes in thickness, flow, and grounding‐line position strongly influence future behaviour. Some recent studies have taken steps in this direction: Jouvet and Cordonnier (2023) focused on the ice‐flow solver component of the simulator and showed that CNNs can approximate higher order ice dynamics, effectively emulating the computationally expensive Glen–Stokes model used to determine ice velocity. Chawla and Holmes (2025) demonstrated how PINNs can model thickness evolution under the shallow ice approximation by approximating the partial differential equations used to describe grounded ice thickness, thereby focusing on the thickness evolution equation component. These studies are promising but focus mainly on the ice dynamics (i.e., ice flow) of small‐scale ice caps and mountain glaciers, rather than on the variability of the entire ice sheet, which includes both ice dynamics and surface mass balance processes. Furthermore, these studies retain certain physical components within their ice‐flow simulations, making them hybrid models. In contrast, our emulator is fully data‐driven, simulating changes across the entire Antarctic ice sheet to provide a more complete, comprehensive, and efficient tool for long‐term Antarctic ice sheet projections. Table 2 compares the three approaches. Physical models are well established and provide high interpretability, but their high computational cost limits their use in ensemble simulations and rapid scenario assessments. Hybrid approaches combine physics with AI models, improving flexibility and efficiency but introducing additional complexity and remaining computationally intensive. Data‐driven approaches use ML to learn models from the data and offer advantages in efficiency and scalability. Once trained, they can generate predictions across large spatial and temporal domains quickly, achieving accuracy comparable to that of traditional models, provided that the training data are representative. While their interpretability is not as straightforward as that of physics‐based approaches, it can be improved using statistical analyses and diagnostic tools. In this work, we propose a data‐driven approach that combines AI's efficiency with explainability methods. Our approach provides transparency in predictions by offering insights into the decision‐making process. We incorporate spatial and temporal features to capture complex ice dynamics, while balancing computational efficiency, accuracy, and explainability. TABLE 2Comparison of physical, data‐driven, and hybrid approaches.ApproachAccuracySpeedInterpretabilityPhysicalHighLowHighHybridHighMediumMedium–highData‐drivenHigh (with representative data)HighMedium© 2026 WILEY‐VCH GmbH Comparison of physical, data‐driven, and hybrid approaches. Methodology The goal of this work is to emulate the computationally intensive numerical PISM (Winkelmann et al. 2011) using efficient ML techniques. Dataset We first use PISM to generate an ice sheet dataset spanning 86 years, from 2015 to 2100 (Golledge et al. 2015). The simulation evolves at a rate of 1.3 model years per processor hour across two processors. For each year, outputs are written on a 51 × 51 grid (2601 grid points per year). At each grid point, we record three climate variables—precipitation, air temperature, and ocean temperature—and three ice sheet properties, ice mask, ice thickness, and ice velocity. Figure 1 shows an example output for the year 2024. We create a tabular dataset in which each grid point for each year corresponds to a row. The task of our ML model is to predict the three ice sheet properties (targets) at each grid point (X, Y) over time, using the climate variables and historical data. Table 1 describes the dataset. Plotted grid data for year 2024. Exploratory Analysis Figure 2 illustrates the distributions of the features and targets for year 2024, in which all exhibit a degree of skewness and non‐normality. The ice velocity distribution shows extreme skewness, with separate plots of values above and below 70. The strong negative correlation between ice thickness and air temperature is depicted in the scatter plot, with its moderate negative correlation with precipitation also illustrated. The negative correlation with precipitation arises from climatic differences in areas of thick (cold, dry East Antarctica) versus thin (warmer, wetter, West Antarctica) ice, rather than any direct causal link. As depicted in the correlation heatmap in Figure 3, the classification of ice mask presents a strong negative correlation with ice thickness. Grid points with thicker ice are more likely to represent grounded ice, whereas those with thinner ice typically correspond to floating ice. Feature distributions for year 2024. Correlation coefficient between variable pairs. Proposed ML Models The goal of our study is to predict annual ice mask, thickness, and velocity measurements using historical data, precipitation, air, and ocean temperatures. Our research presents a solution of explainable spatio‐temporal forecasting models that incorporate exogenous variables to predict multiple target values. We evaluate three ensemble algorithms: Random Forest (Breiman 2001), CatBoost (Prokhorenkova et al. 2018), and XGBoost (Chen and Guestrin 2016). We chose these algorithms for their ability to handle complex, noisy, and skewed data, as well as for mitigating overfitting through techniques such as bagging, boosting, and regularised objective functions. Notably, they can tolerate outliers, extreme values, and zero‐inflated data without relying on data scaling, making them well‐suited to our dataset, which exhibits these characteristics. Random Forest constructs multiple decision trees using bagging from random bootstrap samples, while CatBoost uses gradient boosting on symmetrical trees for faster execution. XGBoost employs a similar gradient boosting framework with a regularised objective function and stands out for its execution speed. All three algorithms can generate feature importance scores, making them valuable for explaining model decisions. Our solution comprises three models: a classification model to predict the ice mask, and two regression models for ice thickness and ice velocity. We propose that all models use the same set of input features. Feature Engineering In addition to using the climate variable features listed in Table 1, we engineer new features to capture the time‐series aspect of the data, the spatial characteristics, and the variable correlations. For the time‐series element, we add values from previous short‐time points (lags) and average values from previous long‐time points to smooth historical data. Similar approaches have been discussed in previous studies (Kämäräinen et al. 2023; Leites et al. 2024; Maddigan and Susnjak 2022). To capture spatial characteristics, we introduce spatial grid features that represent the average values of neighbouring coordinate points within a 3 × 3 grid surrounding each data point. We select only the nearest neighbours to capture spatial correlation at short distances, as it diminishes with increasing distance. In Antarctica, localised transitions near grounding lines and glaciers can exhibit considerable spatial and temporal variability (Miles et al. 2022; Bamber and Aspinall 2013), with notable datasets deriving measurements from the mean of neighbouring cells (Pritchard et al. 2025; Dotto et al. 2021). Therefore, given the low resolution of our dataset, the immediate nearest neighbours are expected to be the most informative. A 3 × 3 neighbourhood also provides a simple and computationally efficient feature that can be generalised across our models. To empirically support this selection, we report the RMSE for larger neighbourhoods. Specifically, we extend the neighbourhood by 2 grid points in each direction yielding a 5 × 5 grid and extend by 3 grid points in each direction yielding a 7 × 7 grid. In addition, we remove the grid feature entirely. Using the climate values (precipitation, air temperature, and ocean temperature) and the historical target values (mask, thickness, and velocity), we create the following engineered features for predicting target values: • Temporal Target Lag Variables: Lag values of 1–4 years for target variables.• Temporal Target Rolling Variables: 10‐year averages for 1‐year lags of target variables.• Temporal Climate Rolling Variables: 10‐year averages of climate variables.• Spatial Target Variables: Average grid values for 1‐year lags of target variables.• Spatial Climate Variables: Average grid values for climate variables. • Temporal Target Lag Variables: Lag values of 1–4 years for target variables. Temporal Target Lag Variables: Lag values of 1–4 years for target variables. • Temporal Target Rolling Variables: 10‐year averages for 1‐year lags of target variables. Temporal Target Rolling Variables: 10‐year averages for 1‐year lags of target variables. • Temporal Climate Rolling Variables: 10‐year averages of climate variables. Temporal Climate Rolling Variables: 10‐year averages of climate variables. • Spatial Target Variables: Average grid values for 1‐year lags of target variables. Spatial Target Variables: Average grid values for 1‐year lags of target variables. • Spatial Climate Variables: Average grid values for climate variables. Spatial Climate Variables: Average grid values for climate variables. Through experimentation, we determined that using 4 years of lag values and 10‐year rolling averages provided sufficient predictive accuracy without increasing feature space or training time. Rolling features require data from the previous 10 years; hence, the prediction horizon of our model spans from 2025 through to 2100, a total of 76 years. Learning Scenarios We explore two learning scenarios: • One‐step ahead forecasting: We update the model annually with historical data for accurate short‐term predictions.• Recursive (multi‐step) forecasting: We use past predictions as inputs to make future predictions, allowing for long‐term forecasting when future data is unavailable. • One‐step ahead forecasting: We update the model annually with historical data for accurate short‐term predictions. One‐step ahead forecasting: We update the model annually with historical data for accurate short‐term predictions. • Recursive (multi‐step) forecasting: We use past predictions as inputs to make future predictions, allowing for long‐term forecasting when future data is unavailable. Recursive (multi‐step) forecasting: We use past predictions as inputs to make future predictions, allowing for long‐term forecasting when future data is unavailable. By investigating both scenarios, we can evaluate the trade‐offs between accuracy and practicality, helping to choose the best forecasting strategy based on data availability and application needs. One‐Step Ahead Forecasting We illustrate the concept of one‐step ahead forecasting in Figure 4. To predict values at each grid point for a given year (e.g., 2025), we train the model on data from previous years (2019–2024). We then update the training dataset to include new data for predicting each subsequent year. In general, we continuously update the training data with past observations from the dataset. We propose two approaches for how much past data to incorporate: Training dataset windows. •Expanding window (Figure 4a): Uses all historical data, but becomes computationally intensive in later years.•Moving window (Figure 4b): Uses only the past 20 years of data, reducing the dataset size and improving efficiency. •Expanding window (Figure 4a): Uses all historical data, but becomes computationally intensive in later years. Expanding window (Figure 4a): Uses all historical data, but becomes computationally intensive in later years. •Moving window (Figure 4b): Uses only the past 20 years of data, reducing the dataset size and improving efficiency. Moving window (Figure 4b): Uses only the past 20 years of data, reducing the dataset size and improving efficiency. Recursive (Multi‐Step) Forecasting In this scenario, we generate forecasts for multiple time periods ahead without updating the training set with actual values. Instead, we use the previously predicted values to retrain the model. For example, when forecasting 2026, we use the predicted 2025 values from the previous cycle. With recursive forecasting, residual errors compound over time. Therefore, we apply both the expanding window and the moving window approaches to our best‐performing algorithm, which is selected based on its one‐step ahead forecasting performance. Both one‐step ahead and recursive forecasting scenarios retrain the models each year, which can be computationally intensive. Thus, we also explore the concept of periodic retraining for our best‐performing algorithm, where models are retrained at fixed intervals to improve efficiency. Our proposed framework is presented in Figure 5. Overview of the proposed modelling framework. The diagram summarises the main stages of the workflow, including input feature generation (precipitation, air and ocean temperatures, and ice sheet variables); spatio‐temporal feature engineering; model training using Random Forest, CatBoost, and XGBoost; and output generation for ice mask, thickness, and velocity predictions across different window types and retraining intervals. Experimental Settings Benchmark Techniques and Parameter Settings We evaluate our models against a naïve benchmark which assumes no change in our target variables over time (a simple random walk) (Armstrong et al. 1984). For hyperparameter tuning, we used a randomised grid search with fivefold cross‐validation for time‐series data using the settings listed in Table 3. However, we found no single set of optimal parameters. The variability in outcomes reflected the algorithms’ stochastic nature and their sensitivity to initial random seeds. The performance metrics were similar across the top‐performing settings, suggesting that the Random Forest model was relatively robust to tuning within the explored ranges. To maintain reproducibility while avoiding overfitting, we use the default settings in Table 3. TABLE 3Hyperparameter settings for Random Forest models.ParameterDescriptionTuning settingsDefaultn_estimatorsThe number of trees100, 200, 300100max_depthMaximum tree depthNone, 10, 20Nonemin_samples_splitMinimum number of samples to split an internal node2, 5, 102min_samples_leafMinimum number of samples at a leaf node1, 2, 41max_leaf_nodesThe maximum number of leaf nodesNone, 300, 400, 500, 800, 1000, 1500, 2000, 3000, 5000None© 2026 WILEY‐VCH GmbH Hyperparameter settings for Random Forest models. Performance Metrics Ice Mask Classification Metrics The precision of a class is the fraction of grid points correctly classified as belonging to the class, out of all the grid points predicted to belong to it. For each ice mask class m∈{2,3,4}, we can define the precision as (1) Precisionm=TPmTPm+FPm where TPm and FPm represent the true positives and the false positives of the class m. The recall of a class is the fraction of grid points predicted to belong to the class out of all the grid points that actually belong to this class. For each ice mask class m∈{2,3,4}, we can define the recall as (2) Recallm=TPmTPm+FNm where FNm represents the false negatives of the class m. The F1 score of a class is the harmonic mean of precision and recall, balancing the two. (3) F1m=2×precision×recallprecision+recall Thickness and Velocity Regression Metrics To evaluate the accuracy of the thickness and velocity model algorithms, we report the root‐mean‐square error (RMSE), a widely used metric of model performance. RMSE provides a measure of the difference between actual and predicted values, with larger deviations being penalised more heavily than with other metrics. The RMSE for the velocity and thickness models, used to predict 2257 grid points each year, can be expressed as (4) RMSE=12257∑i=12257(vi−v^i)2 where vi and v^i represent the actual velocity value and the predicted velocity value, respectively. The Rank Metric To reduce the impact of extreme values on model evaluation, we also rank algorithms by RMSE for each year (Makridakis and Hibon 1995) and then average the ranks across the 76‐year test period. We report both average metrics and ranks (lower is better). Forecast Validity We use the Diebold–Mariano test (Diebold and Mariano 1995) for the thickness and velocity models to determine if each model's predictions are statistically significantly different from the benchmark and competing algorithms. This test evaluates pairwise differences in forecast accuracy at each grid point over 76 years. It calculates the loss differential between the two forecasts and then evaluates a test statistic to check if the expected value of the loss differential is zero. Model Explainability To ensure transparency and trust in our models, we use explainability techniques to provide insights into how our models reach their conclusions. We use two approaches with our best‐performing model: (1) Random Forest's built‐in feature importance 2 based on the Gini index (an impurity‐based measure) and (2) SHAP (based on game theory) (Lundberg and Lee 2017) to obtain both global and local insights. These methods help identify which features drive ice sheet predictions, thereby increasing our understanding of ice dynamics in our models. Results We evaluated each algorithm using one‐step ahead forecasting with an expanding training window from 2025 to 2100, repeating the test four times with different seed initialisations. Additional results for Random Forest models using recursive forecasting and moving training windows are also reported. We summarise the results in Table 4. Ice mask metrics represent the average number of incorrectly classified mask predictions, and thickness and velocity metrics represent the average RMSE. TABLE 4Average metrics for expanding/moving training windows using one‐step and recursive forecasting predicting from 2025 to 2100.ModelAlgorithmWindowOne‐StepRecursiveMetricRankMetricMaskBenchmark naïve—3.503.1831.8CatBoostE3.383.13—XGBoostE3.293.03— Random Forest E 3.19 2.75 30.7 Random ForestM3.242.9130.7ThicknessBenchmark naïve—3.382.3630.6CatBoostE5.644.87—XGBoostE4.203.82— Random Forest E 3.36 1.86 29.6 Random ForestM3.392.1029.6VelocityBenchmark Naïve—67.022.89152.8CatBoostE65.902.83—XGBoostE70.093.55— Random Forest E 65.58 2.78 135.9 Random ForestM64.792.94135.4 Notes: E = expanding training window, M = moving 20‐year training window; Results for the best‐performing models are shown in bold.© 2026 WILEY‐VCH GmbH Average metrics for expanding/moving training windows using one‐step and recursive forecasting predicting from 2025 to 2100. Notes: E = expanding training window, M = moving 20‐year training window; Results for the best‐performing models are shown in bold. One‐Step Ahead Forecasting Our results in Table 4 show that the Random Forest algorithm outperformed other models for all three target variables when using an expanding training window, as highlighted in bold. We note the following observations: •The Random Forest model performed well for ice mask and velocity predictions, with competing models showing similar accuracy. However, it was clearly the best model for ice thickness.•The Random Forest ice mask model improved significantly over the naïve benchmark, misclassifying between 0 and 7 grid points per year out of 2257. On average, it outperformed other models.•For ice thickness, Random Forest had the lowest average RMSE (3.36, range: 1.0–8.6), compared to XGBoost (4.2, range: 2.1–9.1) and CatBoost (5.64, range: 3.54–10.84). While the RMSE gain over the naïve model was small, its ranking was consistently better across 76 yearly test sets.•For ice velocity, Random Forest had an average RMSE of 65.58 (range: 16.9–227.5), similar to CatBoost (65.90, range: 20.3–226.3) and slightly better than XGBoost (70.1, range: 22.1–206.9). It showed clear improvement over the naïve model in both RMSE and ranking.•Gradient boosting models (CatBoost, XGBoost) struggled with zero‐inflated data, sometimes predicting unrealistic negative thickness and velocity values or non‐zero values where zero was expected. •The Random Forest model performed well for ice mask and velocity predictions, with competing models showing similar accuracy. However, it was clearly the best model for ice thickness. The Random Forest model performed well for ice mask and velocity predictions, with competing models showing similar accuracy. However, it was clearly the best model for ice thickness. •The Random Forest ice mask model improved significantly over the naïve benchmark, misclassifying between 0 and 7 grid points per year out of 2257. On average, it outperformed other models. The Random Forest ice mask model improved significantly over the naïve benchmark, misclassifying between 0 and 7 grid points per year out of 2257. On average, it outperformed other models. •For ice thickness, Random Forest had the lowest average RMSE (3.36, range: 1.0–8.6), compared to XGBoost (4.2, range: 2.1–9.1) and CatBoost (5.64, range: 3.54–10.84). While the RMSE gain over the naïve model was small, its ranking was consistently better across 76 yearly test sets. For ice thickness, Random Forest had the lowest average RMSE (3.36, range: 1.0–8.6), compared to XGBoost (4.2, range: 2.1–9.1) and CatBoost (5.64, range: 3.54–10.84). While the RMSE gain over the naïve model was small, its ranking was consistently better across 76 yearly test sets. •For ice velocity, Random Forest had an average RMSE of 65.58 (range: 16.9–227.5), similar to CatBoost (65.90, range: 20.3–226.3) and slightly better than XGBoost (70.1, range: 22.1–206.9). It showed clear improvement over the naïve model in both RMSE and ranking. For ice velocity, Random Forest had an average RMSE of 65.58 (range: 16.9–227.5), similar to CatBoost (65.90, range: 20.3–226.3) and slightly better than XGBoost (70.1, range: 22.1–206.9). It showed clear improvement over the naïve model in both RMSE and ranking. •Gradient boosting models (CatBoost, XGBoost) struggled with zero‐inflated data, sometimes predicting unrealistic negative thickness and velocity values or non‐zero values where zero was expected. Gradient boosting models (CatBoost, XGBoost) struggled with zero‐inflated data, sometimes predicting unrealistic negative thickness and velocity values or non‐zero values where zero was expected. To examine the impact of the neighbourhood grid size, we run our models replacing the 3 × 3 grid features with features representing larger neighbourhoods of 5 × 5 and 7 × 7 grids and also without any grid features. For the mask model, using a 5 × 5 or 7 × 7 grid resulted in higher RMSE values (3.25 and 3.22, respectively) than our chosen 3 × 3 grid (3.19), as well as when no grid features were used (RMSE 3.24). The thickness model with a 3 × 3 grid performed equally well with the 5 × 5 grid (RMSE 3.36) but showed higher RMSE with a 7 × 7 grid and without any grid features (3.41 and 3.39, respectively). The velocity model also showed higher RME values with larger grid sizes (66.48 for 5 × 5 and 65.90 for 7 × 7) and with no grid (65.71), compared to our chosen 3 × 3 grid (RMSE 65.58). These findings support our decision to use a 3 × 3 grid, which aligns with previous literature studies (refer Section 3.4) suggesting that nearest neighbouring cells are the most informative in areas of Antarctica with considerable spatial and temporal variability, particularly given the low resolution of our dataset. We further investigated Random Forest using a 20‐year moving training window, with results also presented in Table 4. Both mask and thickness models showed slight declines in accuracy, but they still outperformed competing models. The velocity model improved slightly in average RMSE, while the average rank was marginally poorer. However, it does not outrank the expanding window model when considering results by year. The improvement in velocity performance with a shorter training horizon may reflect non‐stationarity or concept drift in the underlying climate data, in which recent observations are more relevant for predicting near‐future dynamics. We present additional results for the ice mask classification model in Table 5. Both grounded ice (mask = 2) and open ocean (mask = 4) display excellent precision and recall. The most difficult class to predict, with fewer instances in the data, is floating ice (mask = 3). TABLE 5Average precision, recall, and F1 values for predicting the ice mask using Random Forest.MaskPrecisionRecallF120.9992620.9989090.99908530.9784150.9813330.97987240.9993310.9993800.999355© 2026 WILEY‐VCH GmbH Average precision, recall, and F1 values for predicting the ice mask using Random Forest. Recursive Forecasting The final column in Table 4 reports the Random Forest and Benchmark model metrics for recursive forecasting from 2025 to 2100. As expected, errors compound over time, resulting in larger metrics. Notably, there is little difference in prediction accuracy between a 20‐year moving training window and an expanding window. The Random Forest algorithm outperforms the naïve benchmark for mask, thickness, and velocity predictions. There is a larger relative improvement over the benchmark for the recursive velocity model than for the one‐step ahead approach. Periodic Retraining We examine both learning scenarios under the periodic retraining approach, in which we retrain the model at fixed intervals of 1, 5, 10, or 20 years. Table 6 shows the results with the best performance highlighted in bold. TABLE 6Average metrics for expanding/moving training windows and one‐step/recursive forecasting predicting from 2025 to 2100 retraining models at 1‐, 5‐, 10‐, and 20‐year intervals.ModelAlgorithmWindowOne‐stepRecursive1 year5 years10 years20 years2075*1 year5 year10 year20 year2075*MaskNaïve—3.50N/AN/AN/A3.2731.8N/AN/AN/A21.8RFE 3.18 3.253.293.372.8530.929.329.3 29.3 12.2RFM3.283.413.413.473.1930.929.329.329.313.5ThicknessNaïve—3.38N/AN/AN/A3.2430.6N/AN/AN/A17.7RFE 3.35 3.533.673.993.8429.628.928.9 28.6 15.2RFM3.383.633.794.023.6129.629.029.128.616.8VelocityNaïve—67.02N/AN/AN/A52.10152.8N/AN/AN/A101.2RFE65.3965.9271.0670.7459.48136.3137.6138.8143.7107.3RFM 64.60 67.5574.7770.8563.69 135.5 136.7137.2142.2130.4 Notes: E = expanding training window, M = moving 20‐year training window. 2075* = model predicting from 2075 to 2100 without retraining; Results for the best‐performing models are shown in bold.© 2026 WILEY‐VCH GmbH Average metrics for expanding/moving training windows and one‐step/recursive forecasting predicting from 2025 to 2100 retraining models at 1‐, 5‐, 10‐, and 20‐year intervals. Notes: E = expanding training window, M = moving 20‐year training window. 2075* = model predicting from 2075 to 2100 without retraining; Results for the best‐performing models are shown in bold. •Ice mask: The one‐step ahead mask model with an expanding window and annual retraining retains the best performance, with an average of 3.18 incorrectly predicted cells. All configurations outperformed the naïve Benchmark model (3.50). The recursive model performs well even with less frequent retraining. Its performance dropped only slightly across the 5‐, 10‐, and 20‐year intervals, with all configurations outperforming the benchmark (31.8).•Ice thickness: The one‐step ahead thickness model with an expanding window and annual retraining remains the best‐performing configuration (RMSE: 3.35). The moving window variant retrained annually ranks second, with an RMSE comparable to the naïve Benchmark model (3.38). All remaining configurations performed somewhat inferiorly. However, as observed with the mask model, the recursive thickness model shows a slight decrease in RMSE as retraining intervals increase, with all scenarios outperforming the benchmark (RMSE: 30.6).•Ice velocity: Retraining the one‐step ahead velocity model annually yields better performance than the naïve benchmark. In contrast to the mask and thickness models, the one‐step ahead velocity model showed that a moving window outperformed an expanding window when retrained annually; however, this was not the case for the 5‐, 10‐, or 20‐year retraining intervals. Furthermore, unlike the mask and thickness models, the recursive velocity model's accuracy decreases as the retraining period increases. Nevertheless, its accuracy with a 20‐year retraining interval (RMSE: 143.7) remains superior to the benchmark (RMSE: 152.8). •Ice mask: The one‐step ahead mask model with an expanding window and annual retraining retains the best performance, with an average of 3.18 incorrectly predicted cells. All configurations outperformed the naïve Benchmark model (3.50). The recursive model performs well even with less frequent retraining. Its performance dropped only slightly across the 5‐, 10‐, and 20‐year intervals, with all configurations outperforming the benchmark (31.8). Ice mask: The one‐step ahead mask model with an expanding window and annual retraining retains the best performance, with an average of 3.18 incorrectly predicted cells. All configurations outperformed the naïve Benchmark model (3.50). The recursive model performs well even with less frequent retraining. Its performance dropped only slightly across the 5‐, 10‐, and 20‐year intervals, with all configurations outperforming the benchmark (31.8). •Ice thickness: The one‐step ahead thickness model with an expanding window and annual retraining remains the best‐performing configuration (RMSE: 3.35). The moving window variant retrained annually ranks second, with an RMSE comparable to the naïve Benchmark model (3.38). All remaining configurations performed somewhat inferiorly. However, as observed with the mask model, the recursive thickness model shows a slight decrease in RMSE as retraining intervals increase, with all scenarios outperforming the benchmark (RMSE: 30.6). Ice thickness: The one‐step ahead thickness model with an expanding window and annual retraining remains the best‐performing configuration (RMSE: 3.35). The moving window variant retrained annually ranks second, with an RMSE comparable to the naïve Benchmark model (3.38). All remaining configurations performed somewhat inferiorly. However, as observed with the mask model, the recursive thickness model shows a slight decrease in RMSE as retraining intervals increase, with all scenarios outperforming the benchmark (RMSE: 30.6). •Ice velocity: Retraining the one‐step ahead velocity model annually yields better performance than the naïve benchmark. In contrast to the mask and thickness models, the one‐step ahead velocity model showed that a moving window outperformed an expanding window when retrained annually; however, this was not the case for the 5‐, 10‐, or 20‐year retraining intervals. Furthermore, unlike the mask and thickness models, the recursive velocity model's accuracy decreases as the retraining period increases. Nevertheless, its accuracy with a 20‐year retraining interval (RMSE: 143.7) remains superior to the benchmark (RMSE: 152.8). Ice velocity: Retraining the one‐step ahead velocity model annually yields better performance than the naïve benchmark. In contrast to the mask and thickness models, the one‐step ahead velocity model showed that a moving window outperformed an expanding window when retrained annually; however, this was not the case for the 5‐, 10‐, or 20‐year retraining intervals. Furthermore, unlike the mask and thickness models, the recursive velocity model's accuracy decreases as the retraining period increases. Nevertheless, its accuracy with a 20‐year retraining interval (RMSE: 143.7) remains superior to the benchmark (RMSE: 152.8). We also trained the one‐step ahead and recursive models on data up to 2074 and predicted years 2075–2100 without retraining. The results are shown in the ‘2075’ columns in Table 6. The one‐step ahead mask model continues to outperform the Benchmark model; however, the accuracy of the thickness and velocity models decreases relative to their benchmarks. This may indicate the importance of retraining the model to capture the slow‐moving concept drift present in the data. The results for the recursive method are not directly comparable to the other training regimes because the test set covers a shorter period (2075–2100) than the periodic retraining test set (2025–2100). With fewer test years, the compounded errors in the features are less pronounced. Computational Time Figure 6 compares the computational times for Random Forest (model with the lowest error) and XGBoost (fastest model) using one‐step ahead forecasting and an expanding training window. We include a moving window with Random Forest for comparison. For the mask models, XGBoost and Random Forest perform similarly with an expanding window. However, for the thickness and velocity models, XGBoost with an expanding window outperforms Random Forest with an expanding and moving window. Execution times by forecast year for XGBoost and Random Forest using one‐step ahead models. Figure 7 shows the execution time per year and the cumulative run times for recursive forecasting of the three models combined, i.e., predicting ice mask, thickness, and velocity, using the expanding and moving training windows. In summary, to forecast the 76 years, the expanding training window model requires approximately 5.5 h, the moving training window model about 1.5 h, and the PISM model 58 h (46 min per model year). Cumulative execution times by forecast year for Random Forest using recursive combined model. Diebold–Mariano Significance Test For the thickness and velocity models, we compute the Diebold–Mariano test statistic at each grid point using pairwise comparisons for the expanding window Random Forest model, the moving window Random Forest model, and the naïve Benchmark model. Table 7 shows the number of grid points where the forecast is statistically different between models (at a 5% significance level). The test statistic is not calculated for grid points where the forecasts match exactly. Such instances are recorded in Table 7 as N/A. Often, these are the points in the ocean where we would expect the forecasts of competing algorithms to match. TABLE 7Diebold–Mariano results for pairwise comparisons of the expanding window Random Forest model, the moving window Random Forest model, and the naïve Benchmark model.Naïve vs.Naïve vs.RF (M) vs.RF (E)RF (M)RF (E)ThicknessSig.714777447Non‐sig.338257617N/A120512231193VelocitySig.35015641381Non‐sig.1907693876N/A000© 2026 WILEY‐VCH GmbH Diebold–Mariano results for pairwise comparisons of the expanding window Random Forest model, the moving window Random Forest model, and the naïve Benchmark model. For the thickness target, when comparing the Benchmark model to the Random Forest models, more grid points have significantly different forecasts than non‐significant ones. Hence, we conclude that both Random Forest models provide forecasts that outperform the benchmark. When comparing the two Random Forest models with differing training windows, we observe fewer grid points showing statistically significant differences in forecasts than those considered the same. Therefore, across all grid points, there is no significant difference between the two models. For the velocity target, results show the Random Forest moving window model produces significantly different forecasts from the naïve model, whereas the Random Forest expanding model does not. The moving window model forecasts are significantly different from those of the expanding window, and given its lower RMSE, we infer that the moving window model produces more accurate forecasts. Random Forest Feature Importance We provide insight into the decision‐making process of our Random Forest model with an expanding training window by analysing its feature importance scores. Ice Mask Prediction Figure 8 explains the mask model predictions using the Random Forest impurity‐based feature importance. We calculate variable importance values for each feature and plot the results across the 76 test sets (2025–2100). The top sub‐figure of Figure 8 shows the top five important variables for predicting the ice mask: the previous year's ice thickness, the 10‐year rolling average of ice mask, and ice masks from 1 to 2 years prior. These results explain delayed responses in ice dynamics, where the ice mask from up to 3 years prior has a strong effect because ice flow, accumulation, and melting do not cause immediate changes but unfold over multiple years. The top five variables are engineered features, which indicate that the model benefits significantly from temporal dependencies in the data. Feature importance in predicting ice mask. The importance of input variables is illustrated in the bottom sub‐figure of Figure 8. Air temperature is the largest contributor, with ocean temperature and precipitation appearing similar. The air temperature's direct and immediate effect on surface melting makes short‐term variations more influential than rolling or grid averages. In contrast, rolling features (long‐term averages) are more valuable for ocean temperature and precipitation, capturing long‐term climate trends and lagged effects. Figure 9 illustrates how input features relate to their grid and rolling counterparts in terms of importance. Air temperature has a direct and immediate effect on surface melting, making short‐term variations more influential than grid or long‐term rolling averages. However, for ocean temperature and precipitation, the rolling (long‐term average) features are more valuable in capturing long‐term climate trends and lagged effects. The gradual increase in ocean temperature importance highlights that ocean‐driven basal melting is a slow but persistent process that shapes long‐term ice loss. Snow accumulation and ice formation build up over time, making long‐term precipitation trends matter more for ice sheet mass balance. The importance of each input feature and its rolling values (from previous years) in predicting ice mask. Ice Thickness Figure 10 explains the thickness model predictions using the Random Forest impurity‐based feature importance. The top sub‐figure shows the variable importance values for all features, highlighting a pronounced change near 2030 where the previous year's thickness becomes extremely important. This suggests that ice thickness serves as a key memory effect, encapsulating the cumulative impact of past climate conditions, ice dynamics, and mass balance changes. The irregularity prior to 2030 may be due to the limited training data, causing the model to over‐rely on certain variables. Feature importance in predicting ice thickness. The bottom sub‐figure of Figure 10 focuses on the input variables, showing that ocean temperature is the largest contributor, followed by air temperature and precipitation. This aligns with physical processes, as ocean temperature drives basal melting and ice shelf thinning, which can indirectly influence grounded ice loss. Air temperature directly affects surface melting, while precipitation contributes to accumulation, both of which affect overall ice thickness. The relative weighting of these variables suggests that long‐term ocean‐driven changes have a dominant effect, while atmospheric influences remain crucial in shaping ice dynamics. Ice Velocity Figure 11 explains the velocity model predictions using the Random Forest impurity‐based feature importance. The top sub‐figure shows that the previous year's velocity is extremely important, indicating a strong temporal dependency in ice dynamics. This is because ice flow is a continuous, gradual process with momentum and inertia, making it dependent on previous velocities. Feature importance in predicting ice velocity. The bottom sub‐figure focuses on the input variables, with precipitation being the largest contributor, particularly between 2040 and 2060. Increased snowfall enhances mass accumulation, influencing ice movement through changes in surface slope and gravitational flow. Air temperature initially has a slightly greater influence than ocean temperature. However, ocean temperature becomes more important later, reflecting the delayed but significant effects of basal melting and ice shelf thinning, which can destabilise grounded ice and accelerate velocity. SHAP‐Based Feature Importance Analysis We further enhance the interpretability of our Random Forest model with an expanding training window by applying SHAP analysis. Ice Mask Figure 12 shows SHAP‐based global feature importance for predicting ice mask classes. The bar lengths indicate the impact of each feature on predictions, and the colours represent the distribution of feature importance across ice mask classes. The results show that last year's thickness and recent mask history are the strongest predictors. Their influences vary across mask types: Past mask values dominate grounded ice, recent thickness explains open ocean, and both factors affect floating ice. This suggests that stable regions, such as the open ocean, are driven by current thickness, while transitional zones, such as grounded and floating ice, are more influenced by past ice mask states. Mask model SHAP summary plot. From a local perspective, Figure 13 illustrates model interpretability for the grid point (−486,400, 972,800) near the Filchner Ice Shelf in Figure 14. The SHAP force plot shows that past mask states of 3 (lag and rolling values of grounded ice) strongly increase the likelihood of predicting class 3 while reducing the probabilities of classes 2 and 4. The thicker ice at this location supports this prediction, suggesting that stable ice regions tend to maintain similar mask states. This highlights the persistence of local ice conditions near the Filchner Ice Shelf. Mask model SHAP force plots at grid point (−486,400, 972,800). Grid point (−486,400, 972,800) depicted in red. Ice Thickness Figure 15 shows SHAP‐based feature importance for predicting ice thickness. The most important features are at the top, with points to the right of zero increasing predictions and those to the left reducing them. Red points represent high feature values, and blue points represent low values. Last year's thickness is the most influential predictor, with high values boosting forecasts and low values having a weaker negative effect. Similar but smaller effects are seen in other lagged thickness features, reflecting the persistence of ice conditions. Ocean temperature, air temperature, and precipitation have modest contributions, indicating that the forecasts are driven more by the ice sheet's slow response timescales than by short‐term climate variability. Thickness model SHAP summary plot. Figure 16 explains why the red grid point near the Filchner Ice Shelf is predicted to have a thickness of 2193.74, closely matching the actual value of 2193.79. The force plot shows that the lag1 thickness (2194) has the strongest positive influence, with lag2 thickness (2194) contributing positively to a lesser extent. Conversely, the rolling thickness (2204) has a small negative effect. These combined effects yield a final prediction of 2193.74, driven primarily by recent thickness history, with lag1 and lag2 dominating. Thickness model SHAP force plot at grid point (−486,400, 972,800). Ice Velocity Figure 17 shows SHAP‐based global feature importance for predicting ice velocity. Last year's velocity dominates, with high values strongly increasing predicted velocity and low values slightly reducing it. Other velocity‐related features, such as rolling or lagged values, also contribute, but with smaller effects, suggesting that recent flow history is the key driver of forecasted velocity. Thickness and air temperature follow as secondary influences, with only a modest effect. These results suggest that velocity is mainly shaped by its own persistence, with weaker influences from ice thickness and climate features. Velocity model SHAP summary plot. Figure 18 explains why the red grid point in Figure 14 is predicted to have a velocity of 113.03, reasonably close to the actual value of 106.45. The SHAP force plot shows that both the rolling velocity (79) and lag1 velocity (110.9) have strong positive influences and dominate the prediction. Other velocity‐related features contribute smaller positive effects, leading to the final prediction. This confirms that recent velocity history drives the prediction, while other features play supporting roles. Velocity model SHAP force plot at grid point (−486,400, 972,800). Overall, the SHAP analysis is consistent with the Random Forest feature importance, with both approaches showing that recent thickness and velocity are dominant predictors and identifying climate‐related variables, such as air and ocean temperatures, as secondary influences. Further Analysis of the Difficult Points Figure 19 depicts the location of the ten grid points with the highest residual error, located around the Ross Ice Shelf, the Filchner Ice Shelf, the Jutulstraumen Glacier, and the Amery Ice Shelf. These locations often feature floating ice. Grid points with the highest error depicted in purple. We inspect the data patterns for the point on the Ross Ice Shelf, the largest ice shelf in Antarctica. Figure 20 shows the three climate features and the three target outputs for this point. We observe that precipitation fluctuates, but the underlying pattern remains constant. There is an increasing trend in air temperature. Ocean temperatures show a lower trough near 2060 and a higher peak near 2100. Time‐series plots of grid point (−364,800, −851,200) on the Ross Ice Shelf. We see a change in behaviour near 2075. As ocean temperatures rise and ice velocity increases, ice thickness decreases, and the ice transitions from grounded to floating. The one‐step ahead model predictions, illustrated in green, predict well and closely follow the actual values as we refresh the training data with simulated data. However, the small effect of the input variables is not substantially large enough for the model to predict the change near 2075. Consequently, both recursive models fail to alter their forecasts of ice thickness and mask, thereby compounding errors in future years. This occurs due to weak input signals, error accumulation in recursive forecasting, missing key physical features, and time‐lagged responses in ice dynamics. The recursive model forecasting from 2075 begins on the cusp of the velocity increase, successfully modelling the increasing trend until its decline. Unfortunately, our model lacks features to capture the mechanisms driving this decline, such as ice shelf breakup dynamics, grounding‐line migration, or delayed ice‐ocean interactions, leading to incorrect predictions that continue to increase. Conclusions and Future Work In this study, we developed an explainable ML framework to emulate an example simulation using the PISM, significantly improving computational efficiency while capturing key ice sheet dynamics. Our analysis demonstrated that Random Forests effectively model ice sheet evolution, with past ice thickness, rolling climate features, and spatial neighbourhood information playing crucial roles. By applying impurity‐based feature importance and SHAP‐based analysis, we provided insights into how past ice conditions and long‐term climate trends influence predictions. Given the gradual nature of ice sheet changes, our results show that historical ice conditions remain dominant predictors, as ice mass loss, accumulation, and flow adjust over decades rather than instantaneously. The comparison between one‐step ahead and recursive forecasting further underscores the challenge of accumulated errors over time, emphasising the need for correction mechanisms in long‐term projections. Quantitatively, the Random Forest emulator reproduced ice mask classifications with an average error of only three cells per year, predicted thickness with an average RMSE of 3.36 m, and velocity with an average RMSE of 65.6 m/year across 2257 grid points. These results confirm that our data‐driven approach achieves reliable accuracy while being more than an order of magnitude faster than traditional simulations. This highlights the importance of temporal dependencies, reflecting the inherently slow response of ice sheets to climate forcing. In future, we plan to extend our framework in several directions. We will test the emulator across different ice sheet models to evaluate its consistency and robustness and explore ways to improve generalisability using domain adaptation or transfer learning. We also aim to develop strategies for controlling forecast errors over longer timescales, for example, through adaptive correction schemes that balance accuracy with computational cost. Finally, we plan to build on our feature engineering by using approaches such as Genetic Programming to develop more advanced spatial descriptors that capture directional and long‐range effects, helping approximate processes such as fracture and grounding‐line retreat while maintaining physical consistency. Funding This work was supported by the University Research Fund at Victoria University of Wellington under grant number USRF‐CDRG‐10733, MBIE Endeavour Fund Smart Idea under contract RTVU2445, and MBIE Research Programme under contract RTVU2206 (‘Our changing coast’). Conflicts of Interest The authors declare no conflicts of interest.