Kashif Khan

ML-Assisted Global MPPT under Partial Shading

Developed a reproducible, ML-assisted global MPPT framework for photovoltaic systems operating under partial shading, where the power–voltage characteristic exhibits multiple local optima. A physics-based single-diode PV model with bypass diode activation was used to generate large-scale synthetic data spanning wide irradiance and temperature variations. Multiple regressors (ANN, RF, SVR, GPR, XGBoost, LightGBM, CatBoost) were benchmarked using a scale-invariant target formulation to improve generalization across operating conditions.

The ML-predicted operating point initializes a lightweight, deterministic micro-refinement stage integrated with a boost converter, ensuring convergence to the true GMPP. Extensive dynamic PSC simulations demonstrate >99.95% tracking factor with sub-millisecond inference latency, validated through tracking efficiency, latency analysis, and Wilcoxon significance tests—highlighting suitability for real-time embedded deployment.

Remaining Useful Life (RUL) — NASA C-MAPSS

Developed Transformer+, a lightweight and regime-aware Transformer architecture for turbofan RUL prediction. The model combines local convolutional feature extraction with sensor-wise attention and operating-regime embeddings to handle non-stationary degradation under limited data. Evaluation followed a strict leakage-proof protocol using sliding-window sequences, train-only normalization, and GroupKFold cross-validation at the engine level.

Highlights (cap=100): RMSE of 10.62, 9.35, 7.65, and 10.89 on FD001–FD004, respectively; compact design with ~65K–383K parameters, sub-millisecond inference (≈0.05–0.3 ms), and calibrated uncertainty via Monte Carlo dropout and conformal prediction.

Extrapolation-Aware ML Based Behavioral Modeling for Advanced RF Devices (GaN HEMTs)

Built small-signal S-parameter models for a GaN HEMT across wide conditions such as bias conditions, frequencies and temperatures using XGBoost and Grey Wolf Optimizer (GWO) for hyperparameter search. Ran controlled splits to evaluate interpolation vs. extrapolation along temperature, bias, and frequency. Results: temperature extrapolation was robust; frequency was intermediate; bias was most challenging. GWO-XGBoost consistently reduced MSE/MAE and improved R² over the default model. Validated with Smith charts and 3D response maps at representative bands.

ML-Assisted Microstrip Filter Design (GWO-ANN)

End-to-end EM → ML surrogate → metaheuristic optimization → fabrication pipeline in Advanced Design System (ADS). Designed and validated a high-frequency microstrip filter using an ANN surrogate and Grey Wolf Optimizer for rapid design-space exploration. Hardware verified with Vector Network Analyzer (VNA) measurements.