Optimization of thermionic cooling semiconductor heterostructures with deep learning techniques

We present a deep learning neural network model to find the AlGaAs-based thermionic cooling structures with the best trade-off between the cooling of the lattice and the cooling of the electrons. These devices are based on the electron-phonon interactions, and therefore, the computational requirements to perform the non-equilibrium Green’s function simulations combined with the heat transport and Poisson equations (NEGF+H+P) are very large. The neural network model used is based on the multi-layer perceptron (MLP) machine learning architecture. The comparison between the NEGF+H+P simulations and the values predicted with the MLP gives accurate estimations for the properties studied: gap between the Fermi level of the emitter and the ground state of the quantum well (W), the electron temperature in the quantum well (T
e
), and the cooling power of the lattice (CP). Also, after using the MLP to predict one million of different device configurations we found the heterostructures corresponding to the maximun CP, minimun T
e
, and the best trade-off between both.