Extending the range of graph neural networks with global encodings

Abstract

Abstract Graph Neural Networks (GNNs) are routinely used in molecular physics, social sciences, and economics to model many-body interactions in graph-like systems. However, GNNs are inherently local and can suffer from information flow bottlenecks. This is particularly problematic when modeling large molecular systems, where dispersion forces and local electric field variations drive collective structural changes. We introduce RANGE, a model-agnostic framework that employs an attention-based aggregation-broadcast mechanism that significantly reduces oversquashing effects, and achieves remarkable accuracy in capturing long-range interactions with linear scaling. Notably, RANGE integrates attention with positional encodings and regularization to dynamically expand virtual representations in virtual-node message-passing implementations. Across multiple state-of-the-art baselines, RANGE consistently restores long-range information, enabling the models to correctly predict electrostatic and dispersion-driven behavior even in out-of-distribution extrapolation tasks, where other unmodified baselines fail. Compared with other long-range paradigms, RANGE achieves the highest accuracy while requiring significantly less computational overhead, and it enables stable and scalable molecular dynamic simulations. RANGE offers accurate and efficient modeling of long-range interactions for simulating large molecular systems.