Training Networks with Internal Prestress

We use disordered elastic networks to model amorphous materials. These models often assume that the networks have zero prestress on the bonds so that each bond is at its unstretched length. However many systems, such as biopolymers, have significant internal stresses even in mechanical equilibrium. We are interested in analyzing how this prestress affects the ability of a network to be trained. We found that prestress hinders training because of two causes: (1) prestress hides the local stress information that is typically required for training, (2) removing prestressed bonds during training alters the force balance at each site, thus leading to a change in geometry when a bond is pruned. We will show that the first issue is easily handled by a more general training rule which extracts the necessary stress information for training. However, this protocol is limited to highly coordinated networks and those with low amounts of prestress. The second issue is more difficult to circumvent since it is difficult to evaluate how the alteration of a bond changes the distribution of internal forces. Beyond a limit in the magnitude of the prestress, altering the network geometry overwhelms our ability to train effectively. We analyze quantitatively how much prestress can be in a network before the ability to train is no longer viable.