Interaction networks are introduced as a model capable of reasoning about object interactions and dynamics in complex systems. The model takes graphs as input, performs object- and relation-centric reasoning, and is implemented using deep neural networks. It can simulate physical trajectories, estimate abstract quantities like energy, and generalize to different system configurations. The model is evaluated on physical domains such as n-body problems, rigid-body collisions, and non-rigid dynamics, showing its ability to accurately predict future states and infer abstract properties. The interaction network is the first general-purpose, learnable physics engine, offering a powerful framework for reasoning about objects and relations in various domains. The model combines structured models, simulation, and deep learning, enabling it to handle variable numbers of objects and relations. It is trained to simulate physical systems with thousands of time steps and generalizes to different system sizes. The model's performance is compared to baselines, showing superior accuracy in predicting physical dynamics and estimating energy. The interaction network's architecture allows for efficient learning and inference, with the ability to handle complex, nonlinear interactions. The model's results demonstrate its effectiveness in simulating physical systems and inferring abstract properties, making it a promising approach for AI in reasoning about physical and mechanical systems.Interaction networks are introduced as a model capable of reasoning about object interactions and dynamics in complex systems. The model takes graphs as input, performs object- and relation-centric reasoning, and is implemented using deep neural networks. It can simulate physical trajectories, estimate abstract quantities like energy, and generalize to different system configurations. The model is evaluated on physical domains such as n-body problems, rigid-body collisions, and non-rigid dynamics, showing its ability to accurately predict future states and infer abstract properties. The interaction network is the first general-purpose, learnable physics engine, offering a powerful framework for reasoning about objects and relations in various domains. The model combines structured models, simulation, and deep learning, enabling it to handle variable numbers of objects and relations. It is trained to simulate physical systems with thousands of time steps and generalizes to different system sizes. The model's performance is compared to baselines, showing superior accuracy in predicting physical dynamics and estimating energy. The interaction network's architecture allows for efficient learning and inference, with the ability to handle complex, nonlinear interactions. The model's results demonstrate its effectiveness in simulating physical systems and inferring abstract properties, making it a promising approach for AI in reasoning about physical and mechanical systems.