Despite recent successes, design and development of high-performing deep learning models for scientific data is a time-consuming, trial-and-error-based,manual task that requires both scientific domain and deep learning exper​tise​.To that end, researchers at Argonne have been developing DeepHyper, an open-source software package to automate a few critical tasks in the design of neural networks. DeepHyper tests thousands of different neural network configurations using supercomputers such as ALCF’s Theta, generating better variants until it identifies the best-performing one for a particular task.