Deep Learning framework towards automated experiments and quantum simulations

Datasets with natural images where availability of thousands of samples with large variabilities are present, common deep learning (DL) classification strategies work well. However, for datasets consisting atomically resolved images, networks are required to rapidly adapt to changes to imaging conditions and successfully locate features (nearly identical objects) to make the analyses efficient. The differences in experimental and simulation parameters lead to out-of-distribution drifts. This work introduces DL workflow of ensemble learning and iterative training (ELIT) as an alternative strategy to surpass such challenges. The EL allows for selection of artifact-free features and pixel-wise uncertainty maps by combining multiple networks. The IT part retrains the networks with already realized features, focusing its attention on features present in the data. The features are then used to construct simulation objects to perform first-principles simulations for geometry optimization, property evaluation and temperature-dependent dynamics. Overall, these workflows may be used to better guide experiments while learning from theoretical models.