Machine learning forecasts of laser intensity evolution through underwater turbulence.

This experimental study investigates whether the temporal evolution of received laser-beam intensity can be forecast from intensity images alone under controlled underwater turbulence. Conventional adaptive-optics systems commonly rely on present wavefront measurements and are therefore reactive to measured turbulence-induced distortions. We examine a different prerequisite problem: whether established deep-learning components can be assembled into short-horizon predictors of experimentally measured intensity evolution, without phase measurements or iterative wavefront reconstruction. Synchronized pupil-plane and focal-plane image sequences are recorded after propagation through a Rayleigh-Bénard water-tank turbulence emulator at three turbulence strengths. Two plane-matched forecasting algorithms are evaluated. A latent-space autoencoder-ConvLSTM is applied to the compact, wandering focal-plane intensity pattern, while a 3D spatiotemporal attention network predicts inter-frame intensity changes in the extended pupil-plane distribution. Forecasts are evaluated at horizons of +1, +2, and +3 frames using mean-squared error and structural similarity index, with persistence baselines, regime-specific and all-regime training comparisons, metric-domain analysis, OAM-mode evaluations, and a cross-algorithm diagnostic. The results show that the learned models produce nontrivial short-horizon forecasts beyond persistence, that the two measurement planes favor different algorithmic designs, and that similar intensity-forecasting behavior extends across Gaussian and beam carrying orbital angular momentum (OAM) morphologies while remaining limited to intensity prediction rather than phase structure, modal purity, or topological-charge recovery.