Over the recent years, flapping wing flight has caught attention due to the need for maneuverable, light and green flight propulsion mechanisms and devices. Unsteady flow past flapping wings involves complex non-linear interactions across spatial and temporal scales. These spatio-temporal flow field characteristics can be resolved using high fidelity computational fluid dynamics simulations. However, even such simulations are memory intensive, computationally costly and prohibit real time load predictions and large parametric sweeps that are required for optimal design and control problems. Hence, there is a need for reduced order models that can effectively represent these complex spatio-temporal features in a lower dimensional space, enable future prediction of system dynamics in the lower dimensional space and, interpolate/ extrapolate predictions for unseen parametric instances. The models would thus enable real time prediction, informed exploration of the design space and cut down associated computational and memory costs. Traditional linear projection based reduction methods such as Proper Orthogonal Decomposition (POD) and Dynamic Mode Decomposition (DMD) require a large amount of training data and a large number of modes to capture the significant nonlinear spatio-temporal features of unsteady flow.

With growing hardware and algorithmic advancements, deep neural networks based approaches for nonlinear model order reduction are gaining traction due to their flexibility in implementation, capability to recognise complex patterns, generalise and transfer learn to previously unseen system parameters in training. Autoencoders(AE) are a class of neural network architectures that can capture nonlinear features in the data owing to a nonlinear reduction. However, they require large training data and lack interpretability of the lower dimensional representations. Recent efforts have shown that physics-informed neural networks(PINNs) are capable of encoding hidden invariances and symmetries of dynamical systems to solve forward and inverse problems with limited or no training data. The current study probes the efficacy of autoencoders and PINNs as possible nonlinear alternatives to traditional projection based non-intrusive model reduction approaches. High fidelity simulations obtained for incompressible unsteady flow past a plunging elliptic airfoil in the low Reynolds number regime has been considered for comparison of the proposed approach.