Towards improving decoder performance with machine learned open quantum system simulations

Achieving resource efficient (ie. using a low number of physical qubits), large-scale error-corrected quantum computations will require good models of the noise to be mitigated by quantum error-correcting codes. Current models are mostly based on straightforward assumptions (e.g. depolarizing noise). Classical neural networks have been shown to be capable of abstracting the dynamics of quantum systems.

We learn noise models from open quantum system simulation software. Our approach is an intermediate step until learning noise models from the real quantum hardware will become feasible at scale. Learning from hardware has the following challenges: 1) the noise affecting the quantum computation which is not well understood at scale; 2) the prohibitive cost of running computations on real hardware for extended periods of time required by the classical machine learning methods.

The learned noise models are used for tuning the decoding performance of topological quantum error correcting codes, such as the surface code. We learn noise models for single-, two- and five-qubit operators (ie. plaquettes) and use the models to setup (hyper-)parameters of state of the art decoders (e.g. adapting error detection graph weights). We present encouraging preliminary results.