Predicting circuit success rates with artificial neural networks

Artificial neural networks are powerful tools for modelling highly non-linear functions. When provided with enough data, they can learn otherwise intractable mappings between high-dimensional vector spaces, such as the vector space of 5x515 images, and class or numerical labels. In this work, we leverage this capability by training several state-of-the-art neural network models to predict the success rate of running circuits on several IBM devices. Our networks achieve similar or better accuracy than non-neural network models based on per gate error rates. We also present results from training networks on simulated data generated by non-Markovian error models, a promising future use case.