Artificial Intelligence‐Based Detection of Local Spots in Network Materials Prone to Mechanical Failure

ABSTRACT Predicting the failure of a noncrystalline solid such as silica glass is challenging due to the complex fracture phenomenon. Failure originates from particular local zones where the atomic structure is highly susceptible to local rearrangement. The local yield stress (LYS) method predicts such locations of plastic events during external deformation. Despite the precise prediction accuracy of the location of the failure initiation, the LYS method requires substantial computational resources as it evaluates each local zone separately for rearrangement and is time‐consuming. Owing to new age development in artificial neural networks, a relation can be established between the plastic events and the atomic structure. In this study, a neural network is trained to identify this relation and, hence, predict the locations of atomic rearrangements. This approach results in a heatmap indicating soft spots that highly correlate with elementary events of fracture and allows for reliable incorporation of structural identification to coarser mechanical frameworks that need local information for a physically meaningful description. The advantage of this method is that it can be trained on a specific material and can predict the local soft spots for unseen structural information without the need for multiple time‐consuming molecular simulations.