New AI method combines physics knowledge with reinforcement learning to optimize sensor placement for detecting harmful algae blooms in lakes

PiCSRL uses domain knowledge-based embeddings to improve reinforcement learning for adaptive sensing tasks with limited labeled data

Achieved superior performance on cyanobacterial gene concentration sampling in Lake Erie using NASA PACE hyperspectral imagery (RMSE of 0.153 vs 0.296 for random baseline)

Demonstrated 98.4% bloom detection rate while outperforming traditional methods like Upper Confidence Bound (UCB) algorithm

Addresses the challenge of high-dimensional, low-sample-size datasets by integrating uncertainty-aware physics-informed features into the learning model