Machine learning delta-T noise for temperature bias estimation

Delta-T shot noise is activated in temperature-biased electronic junctions, down to the atomic scale. It is characterized by a quadratic dependence on the temperature difference and a nonlinear relationship with the transmission coefficients of partially opened conduction channels. In this work, we demonstrate that delta-T noise, measured across an ensemble of atomic-scale junctions, can be utilized to estimate the temperature bias in these systems. Our approach employs a supervised machine learning algorithm to train a neural network, with input features being the scaled electrical conductance, the delta-T noise, and the mean temperature. Due to limited experimental data, we generate synthetic datasets, designed to mimic experiments. The neural network, trained on these synthetic data, was subsequently applied to predict temperature biases from experimental datasets. Using performance metrics, we demonstrate that the mean bias-the deviation of predicted temperature differences from their true value-is less than 1 K for junctions with conductance up to 4G0. Our study highlights that, while a single delta-T noise measurement is insufficient for accurately estimating the applied temperature bias due to noise contributions from other sources, averaging over an ensemble of junctions enables predictions within experimental uncertainties. This suggests that machine learning approaches can be utilized for estimation of temperature biases and similarly other stimuli in electronic junctions.