We consider the problem of probe-based quantum thermometry, and show that machine classification can provide reliable estimates over a broad range of scenarios. Our approach is based on the $k$-nearest-neighbor algorithm. Temperature is divided into bins, and the machine trains a predictor based on data from observations at different times (obtained e.g. from computer simulations or other experiments). This yields a predictor, which can then be used to estimate the temperature from new observations. The algorithm is flexible, and works with both populations and coherences. It also allows to incorporate other uncertainties, such as lack of knowledge about the system-probe interaction strength. The proposal is illustrated in the paradigmatic Jaynes-Cummings and Rabi models. In both cases, the mean-squared error is found to decrease monotonically with the number of data points used, showing that the algorithm is asymptotically convergent. This, we argue, is related to the well behaved data structures stemming from thermal phenomena, which indicates that classification may become an experimentally relevant tool for thermometry in the quantum regime.
Download