Predicting the optimal noise strength for solving optimization problems with analog Ising machines

Analog Ising machines are dedicated hardware solvers designed to solve NP hard optimization problems. However, the global optimum is often not found as the system gets stuck in local minima. While several strategies exist to increase the chance of escaping local minima, these methods often require extensive parameter tuning. In this work, we investigate the injection of large noise as a standalone scheme and in combination with annealing to improve the success rate and the time-to-target (TTT) of analog Ising machines for MaxCut problems. We demonstrate that optimizing the noise improves the TTT by several orders of magnitude and renders both approaches competitive with state-of-the-art schemes, such as chaotic amplitude control. Moreover, we are able to predict an optimal noise value based on the problem connectivity, statistics of the coupling matrix and the coupling strength, thereby eliminating the need for costly parameter optimization.