In this work, a Recurrent Neural Network (RNN) model is used to predict the joint center of a one-degree-of-freedom mechanism in 2D rotation from inertial measurement unit (IMU) data. The model is trained using synthetic data and tested using experimental measurements. In half of the trained models, an additional physics-inspired signal is calculated from the IMU data and used to examine whether it can improve prediction accuracy. The results demonstrate that the RNN with physics-inspired feature engineering can effectively learn the relationship between IMU measurements and joint center location, suggesting that neural networks are a promising tool for improving kinematic estimation from sensor data.