Predicting Damage Initiation with Computer Vision

Structural materials are designed to withstand mechanical loading, and are key to ensuring the safety of systems ranging from civil to military. Failure of structural materials can have catestrophic consequences, and so it is necessary to ensure that they perform reliably and predictably in all loading scenarios. In this work, we specifically focus on damage initiation in incipient spall experiments. Work by experimental collaborators at Los Alamos has produced extensive datasets showing void nucleation due to shock loading. Unsurprisingly, it has been observed that grain boundaries (GBs) seem to be preferential sites for void nucleation. Unexpectedly, however, there does not seem to be any definite trend observed in which GBs nucleate voids and which do not; that is, it appears that not all GBs fail equally. Initial analysis has shown that there appears to be no correlation between spall nucleation likelihood and typical GB properties such as energy or CSL ∑ value. Therefore, it is necessary to employ a more sophisticated method of analysis.<br/>Convolutional neural networks (CNN) are a machine learning framework often used in computer vision applications such as image recognition. Multi-channel CNNs (MCCNN) are CNNs that incorporate non-visible channels as well as visible ones. In this work we apply MCCNNs to the problem of void identification in microstructure. Specifically, we train MCCNNs with reconstructed microstructure to resemble the pre-shock state, but categorized by whether the post-shock state contains a void. The channels for the MCCNN include visible channels (micrographs) as well as invisible channels such as GB energy and CSL ∑. In some cases, these properties (such as the ∑ value) are readily available from EBSD analysis. In other cases, they must be computed using a surrogate model. Of particular interest is the GB energy. GB energy is computed automatically using the lattice-matching method, and then used retroactively to estimate the out-of-plane inclination of the boundary. Finally, the stress-state is etsimated automatically using quasi-static elastic simulations in Alamo. By incorporating all of these effects, the model is therefore able to conglomerate physical properties along with geometric properties in order to determine the subtle trends contributing to void nucleation. In this work, the processing and assembly of the training dataset are presented, along with training results from the MCCNN. This framework will eventually be used to rapidly screen, and perhaps even design, optimal damage-resistant microstructures for structural materials with enhanced safety.