Training a Machine Learning Model for Underwater Chemical Source Localization in Simulated Turbulent Flows

Underwater chemical source localization is challenging due to the dynamic and chaotic processes involved. Averaged across long time scales, the geometry of the chemical plume downstream from a source is determined by the mean water flow. At shorter time scales, the turbulence of water tends to create swirls, eddies, and vortices, preventing the observation of a smooth gradient of chemical concentration. Instead, the chemical concentration in the plume downstream from a source is intermittent with mostly near zero concentration interspersed with shorter higher-concentration segments. Underwater vehicles acting as sensor platforms can sample the chemical concentration and measure the water flow as they move along the path through the turbulent plume. Our goal is to process these measurements to predict the chemical source’s location. We propose to train recurrent neural networks to process the input time series jointly consisting of the chemical concentration observations, water flow measurements, and sensor platform movements, and construct a heatmap that represents the probability that the chemical source is located at different locations around the sensing platform. This model is trained based on simulations where the sensor platform moves along different trajectories across numerous scenarios of various source location and strength and water flow characteristics. In each simulated trajectory, the heatmap at each time step is compared to the true source location using a loss function that encourages the model to minimize the expected distance given the heatmap’s source localization predictions and the true source location, which is known during simulation. We measure and report the accuracy of the heatmap prediction under various water flow characteristics for different resolutions. The average angular error for a distant source is 11.25 degrees—promising results for potential real-world application.