DEPS Abstract

Guest >> Sign In

DIRECTED ENERGY PROFESSIONAL SOCIETY

Abstract: 26-DETE-011

UNCLASSIFIED, PUBLIC RELEASE

Feature Engineering for Generalizing Near-Maritime Atmospheric Optical Turbulence Machine Learning Models

As the U.S. Navy integrates high-energy lasers for air defense and communications, atmospheric optical turbulence – quantified by the refractive index structure parameter Cn2 – presents a critical operational challenge by degrading beam intensity. Effective situational awareness requires accurate Cn2 predictions in complex, low-altitude maritime environments. Currently, machine learning (ML) models, though accurate locally, lack the generalizability to function reliably across different geographic regions. Training site-specific models for every global theater is logistically infeasible, so a geographically stable and accurate model is needed.
This project proposes a "Goldilocks" approach: developing a lightweight, generalized ML framework that maintains accuracy across multiple regions. Using three distinct maritime meteorological datasets, the research employs Random Forest and Extra Tree ensemble regression models to balance model interpretability with prediction accuracy. The methodology involves analyzing similarities and differences between these regional datasets, establishing baseline generalizability through cross-testing, and then utilizing physics-inspired feature engineering to enhance model flexibility. By analyzing feature importance and mean squared error (MSE) across varied datasets, this study aims to identify a refinement procedure for creating universal turbulence predictors. Ultimately, this research seeks to bridge the gap between localized ML precision and global operational utility, ensuring the reliability of laser systems in diverse naval environments.

UNCLASSIFIED, PUBLIC RELEASE