Machine-Learning Based Quantum Cascade Laser Design

Conventional methods for QCL design rely on optimization of the active-core layer thicknesses and compositions to achieve desired electron wavefunctions and associated lifetimes. Device performance metrics, such as threshold-current density and overall efficiency, can then be evaluated. This is an iterative and time-consuming process. Human intervention is required to identify relevant electron states, remove wavefunction hybridization, and impose structural constraints. This process often takes hundreds of iterations, and many weeks, before a reasonable design is achieved. The human intervention aspect of this process can be replaced by a data-driven methodology and deep neural networks, which can ultimately be used for inverse device design. Datasets of existing designs and corresponding performance metrics are needed to train the neural network, so that the network can develop intuitive connections between QCL designs and their performances. After training, the network is expected to accomplish a design goal in hours instead of weeks as in the conventional approach.