A Neural Model for High-Performance Scanning Electron Microscopy Image Simulation of Porous Materials

We present a surrogate model in the form of a neural network that can approximately replicate the Monte-Carlo simulations conventionally performed to simulate scanning electron microscopy imaging of porous materials. These materials are of high practical relevance but scanning electron microscopic images of their microstructures feature difficult to interpret artifacts, specifically in areas where the electron beam enters pores in the sample surface and interacts inside the pores. Because of these artifacts, synthetic back-scattered electron and secondary electron images are of high interest both for verifying image interpretation and as training data in a machine learning context. However, the Monte-Carlo simulations of the physical interaction of the electron beam with the solid material are computationally very demanding. Our surrogate model accepts three-dimensional microstructure representations of porous materials in the form of lists of primitives. The system converts these lists to a specific data representation suitable for a neural network. It then uses a convolutional architecture to generate two-dimensional back-scattered electron and secondary electron images in a single forward pass, realizing 4-5 orders of magnitude performance improvement over the first order simulations. Remarkably, the model performs well on arbitrary microstructures like systems of cubes, even though it was trained on structures consisting of spheres and cylinders only.