The paper introduces the *interaction network* (IN), a model designed to reason about objects, relations, and physics in complex systems. IN combines structured models, simulation, and deep learning to perform object- and relation-centric reasoning. The model takes graphs as input, where nodes represent objects and edges represent relations, and it predicts future states and abstract properties such as energy. Evaluations in n-body problems, rigid-body collisions, and non-rigid dynamics show that IN can accurately simulate physical trajectories, estimate abstract quantities, and generalize to different systems. The IN is the first general-purpose, learnable physics engine and a powerful framework for reasoning about complex real-world domains. The paper also discusses related work and provides a detailed implementation and experimental results, demonstrating the model's effectiveness in various physical reasoning tasks.The paper introduces the *interaction network* (IN), a model designed to reason about objects, relations, and physics in complex systems. IN combines structured models, simulation, and deep learning to perform object- and relation-centric reasoning. The model takes graphs as input, where nodes represent objects and edges represent relations, and it predicts future states and abstract properties such as energy. Evaluations in n-body problems, rigid-body collisions, and non-rigid dynamics show that IN can accurately simulate physical trajectories, estimate abstract quantities, and generalize to different systems. The IN is the first general-purpose, learnable physics engine and a powerful framework for reasoning about complex real-world domains. The paper also discusses related work and provides a detailed implementation and experimental results, demonstrating the model's effectiveness in various physical reasoning tasks.