The growing complexity of geological modeling demands intelligent optimization strategies capable of handling nonlinear, multi-modal, and data-scarce environments. This study introduces world, a novel global interaction–inspired metaheuristic algorithm designed for neural network optimization in geoscientific applications. Rooted in the ecological dynamics of interdependent entities, the World algorithm integrates adaptive learning, social influence mechanisms, and environmental feedback loops to achieve dynamic balance between exploration and exploitation. The proposed framework was implemented within an Automated Deep Learning (AutoDL) environment (Auto-Keras) and evaluated across three key geoscientific prediction tasks: lithology classification, shear wave velocity estimation, and Total Organic Carbon (TOC) quantification. Two benchmark datasets—from the North Sea and Horn River Basin—were utilized to ensure geological diversity. Comparative analysis against established optimizers, including Genetic Algorithm (GA), Particle Swarm Optimization (PSO), Bayesian Optimization (BO), and Tree-structured Parzen Estimator (TPE), demonstrated that world consistently achieved faster convergence, higher accuracy, and superior generalization capability. Empirical results revealed an average 12–18% improvement in predictive performance and a 30% reduction in training time, confirming the algorithm’s robustness and computational efficiency. Moreover, its dynamic adaptation strategy significantly mitigated the risk of premature convergence commonly observed in traditional metaheuristics.