Training biases in machine learning for the analytic continuation of quantum many-body Green's functions

We address the problem of analytic continuation imaginary-frequency Green's functions, which is crucial in many-body physics, using machine learning based on a multi-level residual neural network. specifically potential biases that can be introduced due to use artificially created spectral functions are employed train also implement an uncertainty estimation predicted function, Monte Carlo dropout, allows identify frequency regions where prediction might not accurate, and we study effect noise, particular for situations noise level during training different from actual data. Our analysis demonstrates this method indeed achieve high quality prediction, comparable or better than widely used maximum entropy method, but further improvement currently limited by lack true data training. benchmark our approach applying it case SrVO$_3$, accurate function has been obtained dynamical mean-field theory solver works directly real axis.