<p>Machine learning models for predicting mechanical properties in age-hardenable alloys often struggle with experimental datasets characterized by limited diversity with severe homogeneity, wherein aging time varies while composition and temperature remain fixed. To address this, a novel physically constrained noise augmentation strategy was proposed. Unlike standard arbitrary data augmentation, this approach introduces small perturbations strictly bounded by experimental measurement uncertainties to simulate realistic process fluctuations and explicitly break the feature invariance constraint. Using two experimental alloy datasets (age-hardenable Mg-Al-Zn and Al alloys), multiple ML regression models (Extreme Gradient Boosting, Random Forest, Support Vector Regression, Gradient Boosting Regression, and Adaptive Boosting) were trained and evaluated with and without data augmentation. Noise augmentation significantly improves predictive robustness, reducing Mean Absolute Error (MAE) by up to 30%, while yielding SHAP-based feature attributions that better reflect precipitation-hardening mechanisms. Learning-curve and external-validation analyses further indicate that complementing the training set with augmented samples effectively reduces the required experimental dataset size by up to 60%, achieving near-peak performance with 40% of the original data and significantly lowering the trial-and-error effort for alloy optimization. The framework is constrained by incomplete reporting of prior thermomechanical histories such as solution treatment and quenching in the source literature and requires perturbation magnitudes consistent with measurement and process-control tolerances.</p>