Growing concerns about bias and discrimination in automated systems have spurred a surge of research in algorithmic fairness, which aims to design algorithms with formal fairness guarantees. In this work, we focus on fairness in community detection, the task of identifying cohesive subsets of nodes (communities) that are sparsely inter-connected. We assume that nodes are partitioned into groups, based on a sensitive attribute (e.g., gender), and aim for balanced representation of these groups within the detected communities. We propose a novel fair community detection algorithm that builds on the popular Label Propagation method. Our approach draws inspiration from principles in physics to incorporate fairness into the label propagation process. We present experiments on different real and synthetic datasets, where we study the properties of our algorithm and compare with baselines.