Accurately forecasting production in highly heterogeneous reservoirs remains a significant challenge due to the complex, nonlinear interactions and uncertainties introduced by geological variability. Traditional methods, often based on empirical correlations and simplified assumptions, are limited in their ability to capture these dynamics. In contrast, artificial intelligence (AI) and machine learning (ML) offer superior learning capabilities and robustness to noisy and multi-source data, as well as the potential to integrate domain knowledge grounded in reservoir physics. This review provides a comprehensive assessment of recent advances in AI-driven production prediction for heterogeneous reservoirs. It systematically categorizes the impacts of heterogeneity on production behavior and critically compares conventional approaches with AI-based models, demonstrating clear advantages in accuracy, adaptability, and generalization. Special attention is given to modern AI paradigms such as transfer learning, Bayesian inference, self-supervised learning, and ensemble methods, as well as hybrid frameworks that couple physics-informed neural networks with real-time field data. The review also explores emerging frontiers, including federated learning, explainable AI (XAI), and AI-assisted optimization, which enhance interpretability, security, and operational decision-making. Unlike previous reviews that offer broad overviews or narrow algorithmic perspectives, this study provides a criteria-based synthesis, evaluating methods based on predictive performance, computational efficiency, and parameter sensitivity while considering practical engineering constraints. By focusing on state-of-the-art advances and unresolved challenges, this work outlines a forward-looking research agenda for intelligent, interpretable reservoir management and contributes to developing next-generation forecasting strategies that optimize hydrocarbon recovery in geologically complex environments. • AI-driven models outperform conventional methods in predicting heterogeneous reservoirs. • Emerging tools like transfer learning and ensembles improve sparse data forecasting. • Physics-informed and explainable AI enhance model reliability and interpretability. • Future hybrid AI frameworks enable real-time, adaptive reservoir management.
Kavuba et al. (Fri,) studied this question.