Application of Machine Learning and Data Augmentation Algorithms in the Discovery of Metal Hydrides for Hydrogen Storage

Abstract

The development of efficient and sustainable hydrogen storage materials is a key challenge for realizing hydrogen as a clean and flexible energy carrier. Among various options, metal hydrides offer high volumetric storage density and operational safety, yet their application is limited by thermodynamic, kinetic, and compositional constraints. In this work, we investigate the potential of machine learning (ML) to predict key thermodynamic properties—equilibrium plateau pressure, enthalpy, and entropy of hydride formation—based solely on alloy composition using Magpie-generated descriptors. We significantly expand an existing experimental dataset from ~400 to 806 entries and assess the impact of dataset size and data augmentation, using the PADRE algorithm, on model performance. Models including Support Vector Machines and Gradient Boosted Random Forests were trained and optimized via grid search and cross-validation. Results show a marked improvement in predictive accuracy with increased dataset size, while data augmentation benefits are limited to smaller datasets and do not improve accuracy in underrepresented pressure regimes. Furthermore, clustering and cross-validation analyses highlight the limited generalizability of models across different material classes, though high accuracy is achieved when training and testing within a single hydride family (e.g., AB₂). The study demonstrates the viability and limitations of ML for accelerating hydride discovery, emphasizing the importance of dataset diversity and representation for robust property prediction.

1. Introduction

The development of sustainable and efficient hydrogen storage solutions is important in harnessing hydrogen’s potential as bidirectional energy carrier. Various renewable energy sources including solar, wind, and hydroelectric are promising, and hydrogen can be used to store excess electricity in times of over-production. In times of high energy demand, or in cases of emergencies, hydrogen can react chemically or electrochemically with oxygen to produce only water while releasing substantial energy. For instance, 1 kg of hydrogen has the same energy content as 2.4 kg of methane or 2.8 kg of gasoline [1]. Various techniques, including electrolysis, biogas reforming, and photoelectrochemical processes, can be used to produce hydrogen. This versatile energy carrier can be transformed back into power and heat using fuel cells or turbines. Metal hydrides can play a significant role in storing hydrogen, offering a safer and more compact alternative to conventional gas compression and liquid storage methods [2]. These materials chemically bond with hydrogen, allowing for high volumetric storage densities and the advantage of releasing hydrogen upon demand, making them particularly suitable for applications ranging from transportation fuel to power generation [3]. However, the practical application of metal hydrides is hindered by challenges related to their thermodynamic stability, kinetics, and initial activation. Ideally, these materials would absorb and release hydrogen at pressures below 30 bar close to room temperature, exhibit rapid kinetics, minimal hysteresis, negligible need for activation, and resistance to contaminating gases [4]. Moreover, they should avoid the use of critical raw materials and remain cost-effective. Recently, high entropy alloys have extended the range of candidate materials in the field, showing promise due to the large composition space and potential for high hydrogen storage capacities [5,6]. These materials, characterized by a mixture of multiple principal elements, offer a new avenue for creating metal hydrides with tailored properties as the vast compositional space presents both a challenge and an opportunity [7].

Despite their promise, the search for metallic hydrides combining all necessary properties for applications is a daunting task. In this respect, the integration of machine learning (ML) in materials science represents a transformative shift, enabling the acceleration of new materials development through data-driven insights [8,9,10,11]. The dramatic increase in its usage can be attributed to its remarkable ability to capture complex correlations among large datasets and to efficiently predict materials properties. This capability is crucial for discovering novel metal hydrides with optimized hydrogen storage properties, where co-optimizing various properties is essential, as highlighted in recent studies published on this topic [5,12,13,14,15,16,17,18,19]. The first studies showed the feasibility of an ML approach, but had limited predictive power as they used measured properties of the hydrides as predictors in training the models [12,13]. Shortly afterwards, starting from the HYDPARK database and using the Magpie software (https://wolverton.bitbucket.io/, accessed on 15 July 2025), a model trained on features obtained from the materials composition made the approach fully predictive and easy to apply to estimate the equilibrium plateau pressure, enthalpy and entropy of hydrogenation of new hydrides [6]. Some later studies focused on predicting other relevant properties (enthalpy of formation, hydrogen capacity, etc.) for AB₂ hydrides [14,15,16]. A significant number of recent studies focus on high entropy alloys for hydrogen storage [5,17,18,19] using both experimental and Density Functional Theory (DFT) results to generate the necessary datasets for training ML models, which can then be used to navigate this large compositional space to identify alloys with optimal properties. Using DFT generated data can help to overcome the limited availability of experimental data on metallic hydrides; however, some quantities related to the absorption/desorption process are challenging to obtain from theoretical calculations. Hence, the limited size of available datasets for hydrides, often limited to hundreds of items when using experimental data, remains a critical issue.

In this work, we focus on the effect of augmenting the dataset size on ML prediction quality. To this end, we have collected more experimental data published in the literature and nearly doubled the size of the dataset compared to an earlier publication [20]. In parallel, we tested the benefits and limitations of using a data augmentation technique (Pairwise Difference Regression, PADRE) [21] to increase the size of the dataset without using new real experimental data, but rather recombining the original data. Our ML approach is based on predictors generated using the Magpie platform [22,23], which is structure agnostic and based only on the hydride composition. Several models have been tested using a grid search for an extensive hyperparameters optimization. The models were trained to make predictions of the plateau pressure at 25 °C, as well as the enthalpy and entropy of adsorption/desorption, with the goal of refining our previous results. These quantities provide a comprehensive metric of the thermodynamic stability of metallic hydrides and hence can be used to select promising candidates for real applications.

2. Methodology

2.1. Datasets

Initially, we used the same dataset (DS0) as in our earlier publication [20], containing approximately 400 entries for different hydride compositions. For each entry, experimental data for the plateau pressure at 25 °C ($P_{eq}^0$) and its logarithm, the enthalpy (${∆H}_{for}$) and entropy (${∆S}_{for}$) of hydride formation are available. These quantities also allow a complete description of the variation in plateau pressure as a function of temperature (see the first section in Supplementary Materials for further details). These entries are the result of a careful selection carried out in our previous work from over two thousand entries available in the HYDPARK database (details are reported in Supplementary Materials in Ref. [20]). This dataset was used for comparison and to test a few additional models compared to our earlier work. Gathering data from the literature [24], we built an expanded dataset (DS1) containing a total of 806 entries. Most of the added data refer to AB₂ compounds, i.e., Laves phases (C14, C15, C36). In both datasets, we discarded a few entries with ${lnP}_{eq}^0<-20$ pertaining to complex hydrides as their number is too small for ML models to satisfactorily capture the complex correlations between predictors and target properties in this pressure range.

Following our previous approach [20,25], we employed the Magpie software [22,23] to generate a machine learning (ML) dataset from the compositions of the alloys collected in the original (DS0) and expanded dataset (DS1). This process resulted in the creation of 145 features for each entry in the datasets. These features involve a variety of thermo-physical variables, including mean, standard deviation, maximum, minimum, mode, and maximum difference in the elements in the composition formula. More details on the features generated by Magpie can be found in Table S1 in Supplementary Materials and in the original documentation of the software package [22,23].

2.2. Machine Learning

Several ML models (or algorithms) are usually employed and tested to find the best one for a specific dataset. For this study, we experimented with various algorithms, including k-nearest neighbors (kNN), Random Forest (RF), Random Forest with Gradient Boosting (RFGB) and Support Vector Machines (SVM), as implemented in the scikit-learn library [26]. The k-nearest neighbors (kNN) algorithm predicts a data point’s label by looking at the labels of its k closest points in the training set and choosing the majority (for classification) or the average (for regression). A Random Forest algorithm is an ensemble learning method that builds many decision trees on random subsets of the data and features, then combines their outputs to make a final prediction. This reduces overfitting and improves accuracy compared to a single decision tree. While a Random Forest builds many independent decision trees on random subsets of the data and averages their predictions, Random Forest with Gradient Boosting builds trees sequentially, each one correcting the errors of the previous, making them usually more accurate but also more complex and slower to train. A Support Vector Machine (SVM) model for regression predicts continuous values by finding a line (or hyperplane) that fits the data within a specified margin of tolerance, while minimizing errors beyond that margin. These ML models are well known to perform effectively with medium-size datasets, such as DS0 and DS1. In each model, the features used were those obtained from Magpie, whereas the target quantities to predict were experimental values for $P_{eq}^0$, ${∆H}_{for}$ and ${∆S}_{for}$.

Before training, each dataset was converted into a matrix using the Python library Pandas (2.0.3) [27] and randomly shuffled to mitigate potential grouping of similar data. For each ML algorithm, the best values of its hyperparameters were determined to minimize the chosen loss function (mean absolute error, MAE). This optimization was performed using the GridSearchCV function from the scikit-learn library (1.3.2) on a suitable grid of values. The algorithms were trained for each combination of hyperparameter values on the grid, and the corresponding loss function was calculated. Each training iteration utilized 10-fold cross-validation, evaluating the model 10 times with different configurations of train/test sets to obtain more robust statistics on the error. Cross-validation was also used to evaluate the final generalization error using the best values of the hyperparameters determined in the grid search. As performance metrics, we used the mean absolute error (MAE) and the coefficient of determination (R²) both during training and for the final performance evaluation of each model.

In order to check for the existence of hidden patterns, structures, or groupings in our datasets, we applied a couple of well-known unsupervised ML algorithms (k-means++ and DBSCAN (Density-Based Spatial Clustering of Applications with Noise)), as implemented in the scikit-learn library, to test both distance-based and density-based methods. k-means is a clustering algorithm that partitions data into k groups by repeatedly assigning each point to the nearest cluster center and then updating the centers as the mean of their assigned points. It aims to minimize the variance within clusters and maximize separation between them. k-means++ uses an improved initialization method for the k-means clustering algorithm that selects starting centroids more strategically, spreading them out to reduce the chances of poor clustering results. This leads to faster convergence and better overall cluster quality compared to random initialization. DBSCAN is a clustering algorithm that groups together points that are closely packed (high density) while marking points in low-density regions as outliers. Unlike k-means, it does not require specifying the number of clusters in advance and can find arbitrarily shaped clusters. Finally, we note that cluster analysis in machine learning is about grouping data points based on similarity to reveal structure in unlabeled data. This differs from the traditional cluster expansion in physics which refers to a mathematical technique for approximating properties of many-particle systems by systematically expanding interactions in terms of clusters of particles.

For the first method, we determined the best number of clusters (k) by running the algorithm for different values of it (from 2 to 8) and evaluating the results based on the resulting silhouette diagrams. For DBSCAN, the optimal values of the hyperparameters “epsilon” and “min_samples” were also determined according to silhouette diagrams (see the Supplementary Materials for details, Figures S1–S3).

2.3. Data Augmentation (PADRE Algorithm)

PADRE [21] is a data augmentation algorithm that constructs all possible pairs (i, j) between single entries of the training set and test set and for each pair considers the differences in the original features as new ones. Each entry of the augmented dataset (x_augmented) will be formed by concatenating the original features of the pair and the differences, as in Equation (1) [21]:

$$x_{\mathrm{a}\mathrm{u}\mathrm{g}\mathrm{m}\mathrm{e}\mathrm{n}\mathrm{t}\mathrm{e}\mathrm{d}}=x_i\oplus x_j\oplus x_{i^{-}}{x}_j$$

(1)

where x_i and x_j are the features of two entries of the original dataset. Furthermore, the target variable is also modified in the augmented dataset to represent the difference between the pairs. This procedure results in a significant increase in the size of the original dataset (both the number of features and entries) as detailed in

Table 1. Variation in the dimensions of DS0 and DS1 before and after PADRE.

Dataset	Before PADRE	After PADRE
DS0	398 items × 145 features	158.404 items × 435 features
DS1	806 items × 145 features	649.636 items × 435 features

.

These augmented datasets, hereafter referred to as DSA0 and DSA1, were used to train ML models to assess the effectiveness of the data augmentation procedure.

3. Results and Discussion

3.1. Effect of Increasing the Dataset Size with New Real Data

At first, we trained a few additional models with the logarithm of the equilibrium pressure ($ln{P}_{eq}^0$) as the target property and using the original dataset DS0, as in [21], to test possible improvements. For each model, the best combination of hyperparameters was first determined and then used with 10-fold cross-validation to obtain the final MAE. The results are reported in Tables S2–S5 in the Supplementary Materials. While the kNN model showed the worst performance (MAE = 1.99) and will not be considered further, the lowest MAE (1.35) was obtained with the SVM model. This model performed slightly better than the best result in previous work (MAE = 1.52) [21], which was obtained with an RFGB model.

An attempt at retraining an RFGB model in this work led to a similar MAE (minor differences can be attributed to minor changes in the optimal hyperparameters and statistical fluctuations in 10-fold cross-validation results). Despite the slight improvement in the model performance obtained with the SVM model, it is difficult to achieve further significant improvements in predictions by simply testing other ML algorithms or refining the grid search of the optimal hyperparameters.

Hence, in the following, we describe the results obtained with the same ML models when the size of the dataset is increased with both new real data and with data augmentation techniques. First, we compared the distribution of the target pressure considering the old dataset (DS0) and the new dataset (DS1), as illustrated in Figure 1. It is evident that there is a substantial increase in the number of entries in DS1, particularly in the region near $ln{P}_{eq}^0=0$, between −5 and 5. It is also noteworthy that new data are now available in the region above $ln{P}_{eq}^0$ = 5, whereas only a few data points are added below $ln{P}_{eq}^0$ = −5.

As with DS0, we carried out hyperparameter optimization and trained different models on DS1. As before, the RFGB and SVM models show the best performance, leading to MAE values of 1.07 and 1.13, respectively. These values demonstrate a significant improvement compared with the mean errors obtained with DS0 and in our earlier work, clearly showing the benefit of increasing the dataset size. All results are shown in

Table 2. Mean Absolute Error (MAE) and R² values obtained for ML algorithms with $ln{P}_{eq}^0$ as target using the dataset DS1.

Model	MAE	R2
SVM	1.13	0.82
RFGB	1.07	0.84
RF	1.47	0.74
KNN	1.82	0.64

.

The low performance of the kNN model on both the DS0 and DS1 datasets was somewhat expected. Although the model is simple and intuitive, it can struggle with noisy or high-dimensional data; its predictions are often unstable since they depend heavily on local neighbors, and its performance usually lags behind more sophisticated models. In contrast, RF captures nonlinear patterns and interactions, is more robust to noise, and generalizes well even with modest dataset sizes like 400 or 800. Gradient boosting is particularly effective at further improving its performance. Finally, SVM can perform very well if the kernel and parameters are properly tuned, as was done here, especially on smaller datasets such as DS0 and DS1. However, it does not scale well with larger datasets or a high number of features (see the section on data augmentation).

As already pointed out in our earlier work [21], it is important to rationalize the results of ML models with feature importance analysis, which provides an explainable insight into the inner workings of the models and helps understand correlations between the features and the target properties. The results of feature importance were evaluated for the Random Forest model with the methodology proposed by Shapley [28] (SHAP values in Figure 2), which can be applied to all ML models, and with the native method for random forests implemented in the Scikit-learn library (Figures S4 and S5). With both methodologies, the most important descriptor turns out to be the mean_GSvolume_pa (${\nu }_{pa}^{Magpie}$), which is the average ground-state volume of the alloy, calculated from ground-state volumes per atom of the elemental solids (${\nu }_{pa,i}$) as ${\nu }_{pa}^{Magpie}=\sum _i{f}_i{\nu }_{pa,i}$, where $f_i$ are the atomic fractions. This confirms our previous findings (for a detailed analysis and the physical interpretation of this predictor, the reader is referred to our previous work [20]) and is not affected by the increase in size of the dataset. In addition, the other predictors (among the top-ten most important ones) are also similar, though the exact order is slightly different.

3.2. Effect of Increasing the Dataset Size with Data Augmentation

Producing and collecting new data with PCT measurements is time-consuming; hence, it is desirable to enhance the model performance in other ways, if possible. As data augmentation could be helpful to this aim, we present here the results of the application of PADRE method to our datasets. The new datasets DSA0 and DSA1 were generated by applying the PADRE method and then the RFGB model was retrained on both of them, with the same approach adopted for the original datasets (DS0 and DS1). We did not retrain the SVM model due to its poor scalability with dataset size, and we also excluded the kNN and RF models, because of their modest performance in earlier training. Using 10-fold cross-validation on the test set, the average MAE obtained was 1.36 on DSA0 and 1.24 on DSA1. Comparing these results obtained with data augmentation with the previous ones, we can note that the PADRE method led to an improvement in the prediction quality for DS0, but not for DS1. To better understand this outcome, we systematically applied the PADRE method to datasets of different sizes in the range from 100 to 806, i.e., the size of DS1. These datasets were obtained by random sampling DS1, then the PADRE method was applied to each newly created dataset of size 100, 200, 300, etc., and an RFGB model was trained on each one. The results are shown in Figure 3. It is evident that models trained on datasets generated with data augmentation produce better results compared to the original datasets when the size of the latter is less than 500 elements. In contrast, when the original dataset is sufficiently large, data augmentation is less effective and leads to slightly worse results. Moreover, the effect of data augmentation on the MAE value appears to be limited but not negligible on small datasets. The fact that data augmentation improves the quality of prediction on small datasets more than on large ones agrees with the original findings from Ref. [21], though the crossover point (in our case, 500 elements) critically depends on the type of data and features used. The feature importance values evaluated on the augmented datasets show that the most important one is now the difference in ground-state volumes (Figure S4 in the Supplementary Materials), which is consistent with previous results without data augmentation.

Another interesting point to consider is the quality of the predictions across the range of target equilibrium-pressure values. We already pointed out in earlier work (Ref. [20]) that ML models do not generalize well in the wings of the $ln{P}_{eq}^0$ distribution, i.e., at very low/high values, because of the correspondingly low number of samples in the training dataset. To investigate the effect of data augmentation on this issue, we split the data on the dataset DS1 and DSA1 into bins of size 1 in $ln{P}_{eq}^0$ (e.g., −20 to −19, −19 to −18, etc.) and we calculated the MAE of the data in each bin with and without data augmentation using 10-fold cross-validation (Figure 4). The points in Figure 4 represent the MAE for each fold in a certain bin (10 folds for each bin, cross validation has been applied to each bin separately) when training was performed on the dataset DS1 (without data augmentation), whereas the black line represents the average MAE in each bin without augmentation. With respect to our earlier work (Ref. [20]) we can notice a certain overall reduction in MAE values, but still significantly higher values on the wings of the distribution. After applying data augmentation and training on dataset DSA1, MAE values have been evaluated in a similar way for each bin (red squares and red line). It is evident that data augmentation with PADRE does not lead to a decrease in MAE values in any region of the distribution, with a few exceptions in some bins. In some bins at the extremes of the distribution, the application of PADRE generates significantly worse predictions (i.e., higher MAE values). It is unfortunately evident this approach to data augmentation does not improve the accuracy of predictions at high/low pressure values, where data are sparser than in the middle of the range. Hence, the benefits of data augmentation appear to be limited to small datasets and are only marginal compared to the benefits of introducing new real data in the dataset.

3.3. Clustering and Further Analysis

To gain a deeper understanding of dataset DS1, we employed two clustering algorithms, kmeans++ and DBSCAN, from the Scikit-learn library. Both were applied to the entire feature vector and we searched for the optimal number of clusters (k) and other hyperparameters according to the silhouette score in addition to inertia values (more details and the silhouette score and diagrams are reported in Supplementary Materials, Figures S1 and S2). With kmeans++, the optimal number of clusters was found to be k = 4, while with DBSCAN the corresponding number of clusters was found to be k = 8. The clusters determined according to kmeans++ are shown in different colors in Figure 5.

The clustering results are shown in Figure 5, where the target quantity $ln{P}_{eq}^0$ is plotted against the feature ${\nu }_{pa}^{Magpie}$, defined as the average ground-state volume of the alloy calculated from ground-state volumes per atom of the elemental solids (${\nu }_{pa,i}$) as ${\nu }_{pa}^{Magpie}=\sum _i{f}_i{\nu }_{pa,i}$, where $f_i$ is the atomic fraction. This choice is motivated by our earlier and current results, which clearly indicate a correlation between these quantities. However, as Figure 5 shows, the different clusters of materials do not appear to be well separated in this plot. Similar results were obtained with DBSCAN (see Supplementary Materials for details, Figure S3). These limitations on the clusters obtained with respect to $ln{P}_{eq}^0$ vs. ${\nu }_{pa}^{Magpie}$ are also similar to previous results on the DS0 dataset [20] and were already discussed in detail. In the following, we go beyond the current clustering results to conduct a demanding test designed to assess the generalization capability of the trained ML models for hydrides under the most challenging conditions.

It is well-established that it is difficult for an ML model to generalize to new instances (in the present case on new possible hydrides) that differ significantly from the data on which it was trained. To quantify this difficulty in the worst possible case, we used the kmeans++ clustering results to split the dataset DS1 into training and test sets in the following way: the first of the four clusters obtained from kmeans++ is set aside and used as a test set while the other clusters (2, 3, 4) are used for training an RFGB model; after training, the model results are evaluated by calculating the MAE and R² values on the test set (cluster 1) [29]; this step is iteratively repeated choosing each time a different cluster for testing and the others for training (cluster 2 for testing and clusters 1, 3, 4 for training, and so on). In this way, assuming the kmeans++ clustering has successfully split the datasets into homogenous clusters, each time, the test set contains hydrides which are somewhat different from those used in training. The results are shown in

Table 3. MAE and R² values obtained from the cluster reported in column 1 (set aside as test set) when training an RFGB model on all the other clusters obtained from kmeans++.

Kmean++ Cluster Used as Test Set	Number of Instances	MAE	R2
1	531	5.97	−2.68
2	43	2.92	0.1
3	69	3.26	0.01
4	176	1.25	0.82

and are rather unsatisfactory for most clusters except when cluster 4 is used for testing and clusters 1-2-3 for training.

The above results are, however, dependent on how successful the applied clustering algorithms are in splitting the original dataset into homogenous subsets. To further confirm these findings, we used a different clustering of the dataset, based on the material classes defined in the HYDPARK database, i.e., according to different types of hydrides (AB, AB₂, etc.; see Supplementary Materials for more details, Figures S6 and S7). We then repeated the previous approach using these material classes as clusters and the obtained results are reported in

Table 4. Mean Absolute Error (MAE) and R² values obtained from the cluster (material class) reported in column 1 (set aside as test set) when training an RFGB model on all the other clusters (material classes) as defined in the HYDPARK database. Mg refers to Mg-based hydrides, SS refers to solid solutions and MIC refers to intermetallic compounds not included in previous clusters (other classes are self-explanatory).

Material Class	Quantity of Data	MAE	R2
A2B	10	1.65	−4.06
AB	78	3.9	0.10
AB2	454	2.10	0.37
AB5	106	1.80	−0.5
Mg	32	2.68	−0.26
MIC	52	3.74	−0.41
SS	85	2.25	0.54

.

As observed, the results are generally unsatisfactory across most material classes, although for some classes the limited number of instances may significantly affect the reliability of the results. In fact, this is probably the case for A₂B hydrides (only 10 instances in the class) and possibly for Mg and MIC (miscellanea types) with 32 and 52 instances, respectively. However, it does explain the poor results obtained when testing on AB₅ hydrides or even on the largest class AB₂. In this latter case, which is the second best one among all possible classes, the results point out that training with data on hydrides belonging to different material classes but not including AB₂ lead to performances which are significantly lower than those obtained in Section 3.1 Although another factor that may affect the present results is the imbalance between the material classes sizes, it is clear from the results in both Table 3 and Table 4 that the ML models for hydrides cannot be expected to perform well unless they have been trained on datasets containing a relevant number of data on similar hydrides.

On the contrary, models trained on similar hydrides are expected to perform well when applied to make predictions within the same material class. To verify this point, we created a new dataset DS2 containing only hydrides of AB₂ type (the most numerous) for a total of 454 instances. We then retrained both an RFGB and an SVM model using 10-fold cross-validation to evaluate the results as in Section 3.1. The full results for each fold are shown in Table S5 and they show an average MAE = 0.83 and R² = 0.89 for RFGB and MAE = 1.12 R² = 0.72 for SVM, which are significantly lower than the corresponding values obtained on the whole DS1 dataset.

3.4. Validation of Model Predictions

Though the quality of the predictions based on the trained ML models is still limited at extremely low/high pressures, they can capture some trends in the equilibrium plateau pressures in the composition space and can help guide the exploration of new alloy compositions. For example, Figure 6 shows the variation in the predicted plateau equilibrium pressures at room temperature starting from the FeTi binary equimolar phase and partially substituting iron or titanium with other elements (Ni, Mn, Cu, Al, V). In almost all cases, the predicted $ln{P}_{eq}^0$ decreases in agreement with experimental findings [24] with one notable exception when V substitutes Ti, which is also consistent with experiments.

4. Conclusions

This work focuses on investigating the effectiveness and versatility, as well as the limitations, of machine learning (ML) approaches in the design of metal hydrides and in predicting their equilibrium plateau pressure. Building on our previous research, we first checked for possible improvements using different models on the same dataset (DS0) and we found that limited enhancements are possible. This was the case with an SVM model not tested before. However, we found a higher effectiveness of new experimental data in enhancing the performance of the trained models. Specifically, when doubling the size of the dataset (DS1) the improvement in the MAE and R² values of the trained models was significant. Feature importance results on this larger dataset essentially confirm previous findings.

In contrast, the tested data augmentation technique (PADRE) did not enhance model performance, with enhancements being evident only for small original datasets (before data augmentation is applied) and diminishing as the dataset size increased. This possibly points to the fact that, when the quantity of real data is large enough, the improvement that can be obtained by these techniques is limited or null as the model is already capable of capturing the complex correlations present in the data. Furthermore, the application of these techniques does not lead to any improvement in the quality of the predictions at pressure values which are not well represented in the dataset (at high/low pressure regimes), which could represent a significant advantage for certain applications. The choice to test a data augmentation method was motivated by the improvements these techniques have demonstrated on other datasets, particularly in image analysis; however, our results indicate that this may not always be the case. Although other data augmentation methods might prove more successful, the present results make it clear that the effectiveness of real new data is significantly higher.

We also verified that the generalization ability of these trained ML models is somewhat limited to the same type of hydrides which are well represented in the training set. Attempts at predicting properties for hydrides of different classes lead to poor results. Conversely, models trained on well-represented material classes such as AB₂ can perform very well on similar new materials. We emphasize that this limitation is not reflected by the standard metrics commonly adopted in the ML community (e.g., MAE, R², or similar measures), as many ML results are reported solely on the basis of these values using train/test splits or cross-validation. While such metrics are beneficial in that they enable straightforward comparisons across different published studies and provide a uniform way to approach the generalization problem in ML models, they fail to capture the full extent of a model’s generalization ability—particularly in regions of the data space that may be critical for specific applications and materials design.

Despite the above noted limitations, the framework developed here facilitates the interpretation of ML models and can support the rapid screening of new materials with the desired capacity and thermodynamic properties for specific hydrogen storage use cases [3]. Nevertheless, experimental validation remains essential.