Machine-learning-based prediction of a scalar source location from remote sensors in turbulent flow

We employ a machine learning approach to identify the location of a scalar source in turbulent channel flow using remote sensor measurements. The source is positioned inside the channel, while sensors are distributed at specified positions (x_j, y_j, z_j), where j = 1, 2, 3, …, n and n denotes the number of sensors. For each prescribed source location, the resulting passive scalar signals are recorded at the sensor locations. By systematically varying the source position, a dataset comprising sensor signals and corresponding source coordinates is generated. This dataset is used to train a convolutional neural network (CNN), optimized to minimize the root-mean-square error between the predicted and actual source locations. The prediction error, defined as the ratio of the distance between the predicted and actual source locations to the distance between the actual source and the nearest sensor, is generally below 10%, with the highest errors observed when the source is located close to the channel wall. The prediction is performed rapidly, making the method suitable for real-time applications.