A Neural Network Model of an Insect's Wing Hinge Reveals How Steering Muscles Control Flight

Citation

Melis, Johan Matthijs (2023) A Neural Network Model of an Insect's Wing Hinge Reveals How Steering Muscles Control Flight. Dissertation (Ph.D.), California Institute of Technology. doi:10.7907/teej-tb66. https://resolver.caltech.edu/CaltechTHESIS:02272023-213525351

Abstract

The flight system of the fly is remarkable. A fly can execute an escape maneuver in milliseconds, compensate for wing damage when half of the wing is missing, fly in turbulent conditions, and migrate over large distances. While there are many factors that contribute to the robustness and versatility of insect flight, it is the mechanical encoding of wing motion in the wing hinge that allows flies to rapidly and accurately change wing motion over a large dynamic range. The wing hinge consists of several hardened skeletal elements, named sclerites, and a set of twelve steering muscles are attached to some of these components within the exoskeleton. Due to the anatomical complexity and minute size of the sclerites, the way in which the steering muscles alter the mechanical encoding of wing motion in the hinge is poorly understood.

Using genetically encoded calcium indicators and high-speed videography, is is possible to simultaneously image steering muscle activity and wing motion. In order to extract wing pose from the high-speed video frames, an automated tracking algorithm was developed, that used a neural network and model fitting to accurately reconstruct the wing kinematics. The synchronous recordings of wing motion and steering muscle activity were used to train a convolutional neural network that learned to accurately predict the wing kinematics from muscle activity patterns. After training, the convolutional neural network was used to perform virtual experiments, revealing how the steering muscles regulate wing motion. Correlation analysis revealed that the 12 steering muscles have highly correlated activity. The correlation of muscle activity can be approximated well by a 12D-plane, in which all activity has to reside.

To study the function of the sclerites, a bottleneck was introduced in the convolutional neural network. The bottleneck consists of five neurons, or latent parameters, four parameters corresponding to the state of the different sclerites, on which the steering muscles act, and one parameter representing the wingbeat frequency. This so called latent network predicts both the changes in wing motion and muscle activity patterns as a function of sclerite state. The predicted wing motion as a function of sclerite state matches with previous anatomy and electrophysiology studies for the basalare, first axillary and third axillary sclerites. The fourth axillary sclerite has not been studied before, but shows an antagonistic relationship between the hg_1,2 and hg_3,4 muscles, resulting in a strong decrease and increase, respectively, of stroke amplitude, deviation and wing pitch angles.

By replaying the wing kinematics of the virtual experiments on a dynamically scaled robotic fly, a model of the aerodynamic and inertial control forces as a function of steering muscle activity was constructed. This control force model was subsequently integrated in a state-space system of fly flight, which in turn was integrated in a model predictive control simulation that was used to simulate free flight maneuvers. The body motion, steering muscle activity, and wing kinematics of the model predictive control simulations were strikingly similar to the recorded maneuvers of free-flying flies.

The integrative, multi-disciplinary approach that was used to reveal the mechanical logic of the wing hinge, and the control problem that a fly needs to solve to stay airborne, are both unprecedented in prior literature. The methodologies and models of this study will be a valuable resource in future research on how the fly's nervous system controls the complex behavior that is flight.

Thesis Files