In this work, we applied Random Forest (RF), Extreme Gradient Boosting (XGBoost), and Artificial Neural Networks (ANN) to predict key performance characteristics of quantum cascade lasers (QCLs), including material gain, current density, and emission frequency. By developing a machine learning-based surrogate modeling framework that replaces computationally expensive simulations of QCLs, we enable orders-of-magnitude-faster evaluation and optimization of a high-dimensional configuration space. The training dataset was generated using a numerical simulator based on the density-matrix transport model. By combining physics simulations with machine learning, we achieved reliable predictions of device characteristics, with standardized RMSE values ranging from 0.21 to 0.55 for RF, 0.16 to 0.51 for XGBoost, and 0.04 to 0.22 for the ANN model, demonstrating the superior predictive performance of the ANN across all investigated performance characteristics. The ANN was subsequently used to analyze the full configuration space defined by possible layer thicknesses and electric fields. Approximately 44 million configurations were evaluated in about five minutes, achieving a speedup of approximately 90,000 times over the numerical simulator for a single configuration. This approach allowed the identification of designs with improved material gain and facilitated the efficient optimization of key parameters while maintaining high prediction reliability.
In this work, a set of effective semiconductor Maxwell–Bloch equations has been used to study the dynamics of a bound-to-continuum terahertz quantum cascade laser (QCL) under external optical feedback at both threshold and maximum operating currents. The used model is extended to include current-dependent input parameters obtained from density matrix transport simulations, enhancing physical accuracy and enabling device-specific modelling within the relevant current range for a specific fabricated QCL. Near threshold, external feedback shifts the emission from single-mode to multimode and, depending on the external cavity length, can produce different dynamical regimes. The feedback-induced reduction of the threshold current is demonstrated and quantified. At maximum operating current, the laser is inherently multimode, with external feedback producing different dynamical regimes. The model predicts high sensitivity of output power to variations in the external cavity. When operating with a few modes, the output remains interpretable for sensing applications, whereas highly multimode emission leads to noisy, less interpretable signals. This framework provides insight into self-mixing and mode competition phenomena, aiding in the identification of QCL designs that minimise multimode behaviour for reliable THz sensing and has a direct impact on the advancement of QCL-based sensing and communication applications.