Accurately predicting the failure conditions of materials or substructures can be a computationally de- manding task. Especially in composites materials, the large number of possible interactions of multiple damage modes often requires reliance on complex computational models. In this work, we propose a data-driven approach to material damage prediction. In particular, we use support vector machines (SVMs) to learn manifolds that can separate damaged and intact material states and draw a decision boundary between them. Even though such an approach may seem to provide scarce information (namely, a binary signal for damage), we show how the construct of SVM also easily allows to estimate the probability of failure of any given state. In certain instances, this probability can also be related to energy measures used in well-known failure criteria. Furthermore, our work presents different kernels, or similarity measures to learn damage manifolds. In particular, we consider both standard kernels and quantum-computing inspired ones, which have only lately been explored in applications. We show how the architecture of the kernel dramatically influences the learning performance and also propose ways to engineer it, based on domain’s geometry and loading configuration. Preliminary results show that both classically and quantum-computed kernels well predict failure in isotropic materials. In particular, we were able to get practical insight on the manifolds that quantum kernels generate and how these drastically depend on the way that data is encoded. As a very next step, we envision to extend this method to composite structures of increasing complexity, where multiple failure modes can interact and thus generate nontrivial damage patterns.