<p>Changes in the pore pressure of subsurface reservoirs affect their flow performance, geomechanical assessment, storage capacities for various gases, production operations, and risk management. This study introduces a novel approach of using hybrid machine learning (ML) algorithms to predict pore pressure change (u) based on geomechanical inputs from a North Sea (Norway) well. Least squares support vector machines (LSSVM) and multilayer perceptron neural networks (MLPNN) are employed, both individually and in hybrid forms with particle swarm optimization (PSO) and cuckoo optimization algorithms (COA). The available laboratory data reveal strong linear correlations among input variables. To address this and enhance model efficiency, principal component analysis (PCA) was applied to reduce data dimensionality while retaining the majority of the variance in the dataset. The use of PCA helps simplify the input space, reduce the risk of overfitting, and improve computational performance. A scree plot analysis indicated that only two principal components are required to capture most of the data variability and were sufficient for modeling u. The data were split into 70% training (13,126 data points), 10% validation (1,875 data points), and 20% testing (3,751 data points) subsets. Both standalone and hybrid ML models were trained and evaluated five times on these subsets, with the best-performing models selected based on the lowest root mean square error (RMSE) during validation. The results revealed that hybrid models achieved an average RMSE reduction of approximately 70% across the training, validation, and testing phases compared to standalone ML models, demonstrating their superior performance. Among the hybrid models, LSSVM-COA generated the smallest errors, identifying it as the most generalizable model. Further evaluation using scoring and uncertainty analysis confirmed that hybrid models exhibited better performance and lower uncertainty compared to standalone models. The LSSVM-COA model’s robustness was also confirmed by the Williams plot, with 99.58% of the data falling within the applicability domain, indicating minimal influence from outliers. Feature importance analysis of the LSSVM-COA model further revealed that stress–strain decoupling has a greater influence on the model's predictions than the global mechanical loading state. This finding was rigorously validated by Sobol' global sensitivity analysis, which quantified that the stress–strain decoupling principal component (PC2) accounts for 100% of the prediction variance, confirming its dominant physical role in controlling pore pressure changes. This work is among the first to apply hybrid machine learning models, optimized by metaheuristics, directly to large-scale laboratory hydrostatic datasets for pore pressure change prediction enabling robust, generalizable assessment in carbon capture and storage contexts.</p>