Detecting real-time on-chip thermal anomalies is a difficult but a critical problem for designing robust systems. This paper describes a novel approach for real-time anomaly detection of on-chip transient temperature response using an extremely fast prediction model namely the Digital Twin and an Anomaly Detector. The Digital Twin is based on a trained Machine Learning (ML) model to predict the on-chip transient temperature response at specified locations which can be thermal sensor locations on the chip. The Anomaly Detector is an LSTM neural network model which uses the difference between the transient temperature predicted by the Digital Twin and measured data to classify the anomalies in real time.This model classifies the thermal anomalies into several categories. The real time anomaly detection will enable diagnosis of on-chip thermal problems while the chip is operating as part of a system.