Machine-Learning Analysis to Predict the Exciton Valley Polarization Landscape of 2D Semiconductors.

We demonstrate the applicability of employing machine-learning-based analysis to predict the low-temperature exciton valley polarization landscape of monolayer tungsten diselenide (1L-WSe) using position-dependent information extracted from its photoluminescence (PL) spectra at room temperature. We performed low- and room-temperature polarization-resolved PL mapping and used the obtained experimental data to create regression models for the prediction using the Random Forest machine-learning algorithm. The local information extracted from the room-temperature PL spectra and the low-temperature exciton valley polarization was used as the input and output data for the machine-learning process, respectively. The spatial distribution of the exciton valley polarization in a 1L-WSe sample that was not used for the learning of the decision trees was successfully predicted. Furthermore, we numerically obtained the degree of importance for each input variable and demonstrated that this parameter provides helpful information for examining the physics that shape the spatially heterogeneous valley polarization landscape of 1L-WSe.