Training bioinspired sensors to classify flows

Identifying the Origin of Turbulence Using Convolutional Neural Networks

Though turbulence is often thought to have universal behavior regardless of origin, it may be possible to distinguish between the types of turbulence generated by different sources. Prior work in turbulence modeling has shown that the fundamental “constants” of turbulence models are often problem-dependent and need to be calibrated to the desired application. This has resulted in the introduction of machine learning techniques to attempt to apply the general body of turbulence simulations to the modeling of turbulence at the subgrid-scale. This suggests that the inverse is likely also possible: that machine learning can use the properties of turbulence at small scales to identify the nature of the original source and potentially distinguish between different classes of turbulence-generating systems, which is a novel pursuit. We perform numerical simulations of three forms of turbulence—convection, wake, and jet—and then train a convolutional neural network to distinguish between these cases using only a narrow field of view of the velocity field. We find that the network is capable of identifying the correct case with 86% accuracy. This work has implications for distinguishing artificial sources of turbulence from natural ones and aiding in identifying the mechanism of turbulence in nature, permitting more accurate mixing models.

A Direct-Forcing Immersed Boundary Method for Incompressible Flows Based on Physics-Informed Neural Network

The application of physics-informed neural networks (PINNs) to computational fluid dynamics simulations has recently attracted tremendous attention. In the simulations of PINNs, the collocation points are required to conform to the fluid–solid interface on which no-slip boundary condition is enforced. Here, a novel PINN that incorporates the direct-forcing immersed boundary (IB) method is developed. In the proposed IB-PINN, the boundary conforming requirement in arranging the collocation points is eliminated. Instead, velocity penalties at some marker points are added to the loss function to enforce no-slip condition at the fluid–solid interface. In addition, force penalties at some collocation points are also added to the loss function to ensure compact distribution of the volume force. The effectiveness of IB-PINN in solving incompressible Navier–Stokes equations is demonstrated through the simulation of laminar flow past a circular cylinder that is placed in a channel. The solution obtained using the IB-PINN is compared with two reference solutions obtained using a conventional mesh-based IB method and an ordinary body-fitted grid method. The comparison indicates that the three solutions are in excellent agreement with each other. The influences of some parameters, such as weights for different loss components, numbers of collocation and marker points, hyperparameters in the neural network, etc., on the performance of IB-PINN are also studied. In addition, a transfer learning experiment is conducted on solving Navier–Stokes equations with different Reynolds numbers.

Deep learning model inspired by lateral line system for underwater object detection

Inspired by the lateral line systems of various aquatic organisms that are capable of hydrodynamic imaging using ambient flow information, this study develops a deep learning-based object localization model that can detect the location of objects using flow information measured from a moving sensor array. In numerical simulations with the assumption of a potential flow, a two-dimensional hydrofoil navigates around four stationary cylinders in a uniform flow and obtains two types of sensory data during a simulation, namely flow velocity and pressure, from an array of sensors located on the surface of the hydrofoil. Several neural network models are constructed using the flow velocity and pressure data, and these are used to detect the positions of the hydrofoil and surrounding objects. The model based on a long short-term memory network, which is capable of learning order dependence in sequence prediction problems, outperforms the other models. The number of sensors is then optimized using feature selection techniques. This sensor optimization leads to a new object localization model that achieves impressive accuracy in predicting the locations of the hydrofoil and objects with only 40% of the sensors used in the original model.

Learning efficient navigation in vortical flow fields

Efficient point-to-point navigation in the presence of a background flow field is important for robotic applications such as ocean surveying. In such applications, robots may only have knowledge of their immediate surroundings or be faced with time-varying currents, which limits the use of optimal control techniques. Here, we apply a recently introduced Reinforcement Learning algorithm to discover time-efficient navigation policies to steer a fixed-speed swimmer through unsteady two-dimensional flow fields. The algorithm entails inputting environmental cues into a deep neural network that determines the swimmer's actions, and deploying Remember and Forget Experience Replay. We find that the resulting swimmers successfully exploit the background flow to reach the target, but that this success depends on the sensed environmental cue. Surprisingly, a velocity sensing approach significantly outperformed a bio-mimetic vorticity sensing approach, and achieved a near 100% success rate in reaching the target locations while approaching the time-efficiency of optimal navigation trajectories.

Learning hydrodynamic signatures through proprioceptive sensing by bioinspired swimmers

Objects moving in water or stationary objects in streams create a vortex wake. Such vortex wakes encode information about the objects and the flow conditions. Underwater robots that often function with constrained sensing capabilities can benefit from extracting this information from vortex wakes. Many species of fish do exactly this, by sensing flow features using their lateral lines as part of their multimodal sensing. To replicate such capabilities in robots, significant research has been devoted to developing artificial lateral line sensors that can be placed on the surface of a robot to detect pressure and velocity gradients. We advance an alternative view of embodied sensing in this paper; the kinematics of a swimmer's body in response to the hydrodynamic forcing by the vortex wake can encode information about the vortex wake. Here we show that using artificial neural networks that take the angular velocity of the body as input, fish-like swimmers can be trained to label vortex wakes which are hydrodynamic signatures of other moving bodies and thus acquire a capability to 'blindly' identify them.