Autonomy and electrification have emerged as key drivers of innovation and growth in the automotive industry in recent years. Various advanced driver assistance systems (ADAS) have been developed as OEMs race towards a fully autonomous and zero-emissions future. Radar-based ADAS such as adaptive cruise control (ACC), automatic emergency braking (AEB), blind spot detection (BSD), among others are currently being implemented in today’s vehicle. Connectivity related ADAS such as vehicle-to-vehicle (V2V) and vehicle-to-everything (V2X) are poised to become ubiquitous in coming years. Before advanced driver assistance systems are deployed, they need to go through rigorous testing and validation. While measurement is valuable, simulation has emerged as the more practical, cost effective and safer approach of testing the robustness of these systems, especially in corner cases. Here, we present high fidelity physics-based electromagnetics simulations for V2V and automotive radar systems. We will demonstrate the impact of vehicular obstructions, multipath effects, and non-line-of-sight (NLOS) conditions on 5.9 GHz designated short-range communication (DSRC) links. We will also compare simulation results to measured results. Automotive radar simulations of a 128-channel, 77-GHz, multiple input multiple output (MIMO) sensor for azimuthal angle-of-arrival determination will also be presented. Using this virtual sensor, we will demonstrate how simulation can be used to predict sensor performance in low and high-density traffic regions. This MIMO framework is extended to a 512-channel, 4D radar. Here, we will present scatter plots of 4D radar-returns overlaid on different full-scale traffic scenes. Recent research efforts have focused on using radar to classify vulnerable road users (VRUs). Specifically, variations in the Micro-Doppler response of actors in road traffic scenarios can be exploited to classify them. Crucial to this concept is the massive quantity of spectrograms required for training convolutional neural networks (CNNs). Simulation has emerged as the economically viable and practical solution for collecting and accurately labeling these spectrograms. Here, we will present a workflow for obtaining large volumes of spectrograms using high-fidelity, physics-based simulation techniques. Spectrograms obtained from this effort are used to train a 5-layer CNN (see Fig. 1) to classify 4 target classes (car, cyclist, pedestrian, and 4-legged-animals). Results from this study demonstrate nearly 100% classification accuracy after 10 epochs. We also present results from a study of the impact of data quantity used in training and the radar sensor pulse-repetition-frequency (PRF) on the CNN classification accuracy.