Neural Network-Based Intuitive Physics for Non-Inertial Reference Frames

Classical mechanics offers us reliable means to predict various physical quantities, but it is difficult to derive the precise dynamic equations underlying most phenomena and obtain physical quantities in real-world situations. Intuitive physics, the ability to intuitively understand and predict physical phenomena, prevents this complication. However, its applications are confined to the inertial frame of reference. Here, we explored the potentials of neural network-based intuitive physics for solving non-inertial reference frames. We designed three experiments, each of which represents different types of real-world challenges. The task required predicting the speed of an object while the observer accelerates. We demonstrated that multilayer perceptron, invariant methods, and long-term memory networks successfully learn underlying dynamics from observations. This implies that neural network-based intuitive physics provides alternative means to predict various quantities in real-world applications that are unsolvable by classical physics methods.