Predicting drill penetration rate (PR) in rock environments remains a significant challenge due to the complex interplay between rock texture, drilling fluid properties, and operational parameters. Traditional empirical models often lack generalizability and are based on inconsistent datasets, limiting their reliability. To address these limitations, this study develops a comprehensive experimental dataset using rock samples collected from various mines in Iran, tested under controlled laboratory conditions with different drilling fluids, bit loads, and rotational speeds. Texture coefficient (TC), electrical conductivity (EC), load on bit (LOB), and bit rotational velocity (BRV) were selected as input features. Four machine learning models—support vector regression (SVR), stochastic gradient descent (SGD), K-nearest neighbors (KNN), and decision tree (DT)—were trained to predict PR. A biogeography-based optimization (BBO) algorithm was employed to fine-tune hyperparameters and enhance model accuracy. Additionally, a novel hybrid error index (HEI) was introduced to comprehensively evaluate model performance. Among all models, the DT achieved the best accuracy with an HEI of 0.3753, followed by KNN, SVR, and SGD. These findings demonstrate the potential of the DT model, combined with optimized learning and a robust dataset, to reliably predict penetration rate in rock-based engineering projects.