Materials Data Science and Informatics (IAS-9)

For electron microscopy (EM) of beam sensitive materials, damage induced by the probing beam is a common problem. This is illustrated by an HAADF-STEM image at 200 kV acceleration voltage. With an increasing electron dose from left to right, an increasing damage is shown, not only by changes in the Electron Energy Loss Spectra (insets) but also by contrast changes in the images. Hence, reducing beam damage is relevant to many fields in EM, such as e.g., grain boundary and dislocation imaging.

Beam damage can be mitigated by, e.g. increasing the scanning step size. However, this leads to a lower image resolution. Since this represents a constraint for the achievable image quality, it is of interest to overcome this drawback by computational means.

With the onset of deep learning (DL), resolution enhancement has become a computationally feasible task in several areas of image acquisition. The overall procedure is to train a neural network with a ground truth set of matching low- and high-resolution images and then use this network to infer the missing information details of low-resolution images taken at an according lower beam exposure.

As a first approach, a generative adversarial network (GAN) was applied to enhance atomically resolved HAADF-STEM images of SRO. As a benchmark test, we first artificially reduced the image resolution by a factor of 2 and then tried to increase it again by applying the GAN. A first result is shown in Fig. 3, where the same ROI is shown respectively, in artificially lowered, enhanced high and originally high resolution. The enlarged insets clearly show that the GAN resolution enhancement is able to successfully re-produce specific details from the high-resolution image.