To improve modeling of orthotropic structural composites a new failure criterion has been created using multiscale modeling. Repeating Unit Cells (RUCs) and Representative Volume Elements (RVEs) are used to model the constituent parts of the composite. The RVE is then subjected to multi-axial state of stress and the first instance of element failure in the RVE is detected and tagged as a point in the failure point cloud data (FPCD) in the stress/strain space. To use the generated data during an actual finite element analysis requires that accurate failure onset predictions be made when the state of stress in a finite element is compared against the generated FPCD. In this paper, neural networks (NNs) are used to generate an accurate, robust, and efficient predictive model using the generated FPCD. First, the model is evaluated independently using synthetic stress state data. The NN approach is compared against two techniques—Approximate Nearest neighbor (ANN) and Simplified Approximate Nearest Neighbor (SANN). In terms of both accuracy and speed, the NN approach shows superior results. Next, the NN approach is incorporated in a commercial finite element program and evaluated using multi-element finite element models. Results indicate that the NN predictions have promise as the predictions are accurate and efficient.