Atrial arrhythmias are the most frequent rhythm disorders in humans and often lead to severe complications such as heart failure and stroke. While different mapping techniques have provided significant information on the electrophysiological processes associated with atrial fibrillation (AF), the mechanisms underlying AF initiation and maintenance remain unclear. Hence the treatment of atrial arrhythmias is still based on empirical considerations. To assist the study of the complex spatio-temporal dynamics of AF, a realistic-size computer model of human atria was developed. The model geometry was derived from magnetic resonance images of the human heart. Mathematical models of cell electrophysiology describing the ionic currents through the cell membrane were used. By representing the domain as a three-dimensional monolayer, the computational load was sufficiently reduced to allow the simulation of more than 20 seconds of an arrhythmia. With this model, simulated AF, i.e. multiple reentrant wavelets, were induced using different clinically relevant protocols. The model outputs both transmembrane potential maps and electrograms at any location in the atria, facilitating comparisons of simulation results to experimental or clinical data. It is also possible to study separately the conditions leading to the initiation and perpetuation of AF, and, more generally, to uncouple the phenomena by controlling separately the parameters affecting the simulation. First, the mechanisms leading to AF initiation and perpetuation were investigated. In a model of normal conduction in the atria, electrically-induced AF was unsustained and converted to sinus rhythm after a few seconds. After remodeling (applied as an abrupt alteration of tissue properties), however, episodes of sustained AF were obtained. Simulated AF was observed as several wavelets propagating randomly over the whole atrial surface and undergoing anatomical and functional reentries, collisions, and annihilation by mutual interaction. The simulation studies suggest that the restitution dynamics (describing the dependence of the action potential duration on the previous diastolic interval) and the wavelength (effective refractory period × conduction velocity) play a crucial role in determining the duration of AF. Electrograms were then computed during simulated AF and their morphology was characterized by their amplitude and asymmetry. These simulated electrograms were similar to those recorded in humans. By simulating wavefront propagation in carefully prepared conditions, is was possible to determine the effect of the different components of AF dynamics (wavefront shape, collisions, conduction blocks, wavelength) as well as the influence of the underlying substrate (tissue conductivity, anisotropy, heterogeneity) on waveform morphology. Analysis of the amplitude and symmetry of unipolar atrial electrograms is believed to provide information about the electrophysiological substrate maintaining AF.