Accurate runway friction measurements are crucial for aviation safety and operational efficiency, yet traditional assessment methods face limitations in timeliness and consistency. This thesis investigates the potential of machine learning (ML) to address this challenge by evaluating a hybrid model combining eXtreme Gradient Boosting (XGBoost) and Long Short-Term Memory (LSTM) networks. Using a historical dataset containing over five years of environmental measurements and friction measurements provided by Sarsys ASFT, we developed and trained the hybrid model, alongside standalone LSTM and XGBoost baselines, following data preprocessing and hyperparameter tuning.

Our research shows that despite using theoretically sound ML architectures, all models ended up with consistently poor predictive performance. The tuned hybrid LSTM-XGBoost model achieved an R-squared score of -0.3049 on the test set, performing worse than a simple mean-prediction baseline.

This study concludes that, with the provided data, it is not possible to achieve an accurate Hybrid LSTM-XGBoost model for runway friction predictions. The limitations inherent in the data, such as sparsity and inconsistencies, fundamentally hinder the model's ability to learn and generalize effectively.