This article explores the use of machine learning (ML) models, specifically Support Vector Regression (SVR) and Gaussian Process Regression (GPR), to predict nuclear mass excesses using experimental data and a physics-based feature space. The study aims to assess the performance of these models in predicting nuclear mass excesses, both within the training and test regions and beyond, where experimental data is limited. The models are trained on data from the Atomic Mass Evaluation 2020 (AME2020), which includes 2386 nuclei with Z, N ≥ 8. The data is divided into training (75%) and test (25%) sets, with an additional set of 71 newly measured nuclei used for extrapolation. The feature space used in the models includes 12 inputs such as Z, N, A, A^{2/3}, (N-Z)/A, Z_{eo}, N_{eo}, ν_Z, ν_N, PF, Z_{shell}, and N_{shell}. These features capture various nuclear properties and are essential for accurate predictions. The models are evaluated based on their root mean square (rms) errors, with GPR-12 achieving an rms error of 0.14 MeV for the training set and 0.26 MeV for the test set, outperforming many existing mic-mac mass models. The study also examines the extrapolation capabilities of the models beyond the training and test regions. The results show that GPR models perform better in predicting nuclear mass excesses, especially for medium-heavy and heavy nuclei, while SVR models require more data for accurate predictions. The models are further evaluated using the Garvey-Kelson (GK) mass relations, which are based on the independent particle shell model. The results indicate that GPR models maintain these relations well within the training and test regions but show deviations in the extrapolation region. Additionally, the study incorporates Explainable AI (XAI) techniques, specifically SHapley Additive exPlanations (SHAP), to interpret the model's predictions and understand the contributions of individual features. The SHAP analysis reveals that features such as A^{2/3}, Z, A, and N are the most significant in predicting nuclear mass excesses. The study concludes that both SVR and GPR models are effective and reliable tools for predicting nuclear mass excesses, with GPR models showing superior performance in handling a diverse range of nuclear data. The results highlight the potential of these ML models in nuclear physics and suggest that further refinement could enhance their performance, particularly near the drip lines.This article explores the use of machine learning (ML) models, specifically Support Vector Regression (SVR) and Gaussian Process Regression (GPR), to predict nuclear mass excesses using experimental data and a physics-based feature space. The study aims to assess the performance of these models in predicting nuclear mass excesses, both within the training and test regions and beyond, where experimental data is limited. The models are trained on data from the Atomic Mass Evaluation 2020 (AME2020), which includes 2386 nuclei with Z, N ≥ 8. The data is divided into training (75%) and test (25%) sets, with an additional set of 71 newly measured nuclei used for extrapolation. The feature space used in the models includes 12 inputs such as Z, N, A, A^{2/3}, (N-Z)/A, Z_{eo}, N_{eo}, ν_Z, ν_N, PF, Z_{shell}, and N_{shell}. These features capture various nuclear properties and are essential for accurate predictions. The models are evaluated based on their root mean square (rms) errors, with GPR-12 achieving an rms error of 0.14 MeV for the training set and 0.26 MeV for the test set, outperforming many existing mic-mac mass models. The study also examines the extrapolation capabilities of the models beyond the training and test regions. The results show that GPR models perform better in predicting nuclear mass excesses, especially for medium-heavy and heavy nuclei, while SVR models require more data for accurate predictions. The models are further evaluated using the Garvey-Kelson (GK) mass relations, which are based on the independent particle shell model. The results indicate that GPR models maintain these relations well within the training and test regions but show deviations in the extrapolation region. Additionally, the study incorporates Explainable AI (XAI) techniques, specifically SHapley Additive exPlanations (SHAP), to interpret the model's predictions and understand the contributions of individual features. The SHAP analysis reveals that features such as A^{2/3}, Z, A, and N are the most significant in predicting nuclear mass excesses. The study concludes that both SVR and GPR models are effective and reliable tools for predicting nuclear mass excesses, with GPR models showing superior performance in handling a diverse range of nuclear data. The results highlight the potential of these ML models in nuclear physics and suggest that further refinement could enhance their performance, particularly near the drip lines.