AI-Enhanced Automatic Design of Efficient Underwater Gliders

TL;DR

This work tackles the limited hull-shape diversity of underwater gliders by introducing an AI-enhanced automated design framework. It couples a reduced-order deformation cage for expressive yet compact hull representations with a differentiable neural-fluid surrogate to predict hydrodynamics, enabling end-to-end co-optimization of shape and control via CMA-ES across multiple angles of attack. The approach yields fabrication-ready hull designs that, when manufactured and tested, surpass traditional torpedo-like gliders in energy efficiency, as demonstrated by wind-tunnel validation and underwater experiments, and is supported by dynamic simulations and modular hardware design. The resulting pipeline reduces development time for novel, non-trivial glider geometries and has significant potential for long-range ocean exploration and environmental monitoring, with future work focusing on thinner shapes, improved maneuverability, and tighter simulation-to-reality integration.

Abstract

The development of novel autonomous underwater gliders has been hindered by limited shape diversity, primarily due to the reliance on traditional design tools that depend heavily on manual trial and error. Building an automated design framework is challenging due to the complexities of representing glider shapes and the high computational costs associated with modeling complex solid-fluid interactions. In this work, we introduce an AI-enhanced automated computational framework designed to overcome these limitations by enabling the creation of underwater robots with non-trivial hull shapes. Our approach involves an algorithm that co-optimizes both shape and control signals, utilizing a reduced-order geometry representation and a differentiable neural-network-based fluid surrogate model. This end-to-end design workflow facilitates rapid iteration and evaluation of hydrodynamic performance, leading to the discovery of optimal and complex hull shapes across various control settings. We validate our method through wind tunnel experiments and swimming pool gliding tests, demonstrating that our computationally designed gliders surpass manually designed counterparts in terms of energy efficiency. By addressing challenges in efficient shape representation and neural fluid surrogate models, our work paves the way for the development of highly efficient underwater gliders, with implications for long-range ocean exploration and environmental monitoring.

Figures (6)

• Figure 1: Efficient underwater gliders. (a) and (b) Efficient underwater glider designs differ based on the angle of attack, and our algorithm discovers a span of them. We illustrate 4 representative optimal designs for different angles of attack (AoA) and their corresponding efficiency versus AoA curve. (c) Two glider designs were chosen for fabrication and tested as modular outer shells for an internal hardware assembly. (d) We showcase the fabricated designs navigating underwater and the trajectory of the double-wing underwater glider during one dive cycle.
• Figure 2: Computational design framework. Our co-design framework computes both optimal shape and control for the underwater glider. Shape and control serve as input to our efficient neural fluid model, which efficiently computes the hydrodynamic parameters of the glider. Leveraging these parameters, our framework accurately simulates the performances of these gliders. We then leverage an optimization framework that computationally optimizes their performances (i.e., the lift-to-drag ratio $\eta$), yielding an optimized shape for various angles of attack. On the right, we display two optimal designs for the 9 degrees and 30 degrees.
• Figure 3: Efficient computational fluid dynamics. A key component of our design framework is a neural network-based fluid surrogate model. (Top) Dataset: We begin with a set of base shapes, which are represented using a cage representation. This approach allows for efficient interpolation between shapes. (Middle) Training: Instead of relying on traditional fluid simulators, we train a neural surrogate model to predict drag and lift coefficients based on shape and control parameters. (Bottom) Validation: We validate our neural surrogate model using a testing dataset and further confirm its accuracy through wind tunnel experiments. Finally, we validate the overall design by testing the glider's performance in a swimming pool.
• Figure 4: Dynamics Simulation. Our dynamics modeling tools accurately simulate the glider's transition from static modes to downward gliding modes.
• Figure 5: Hardware overview. (a) The internal hardware tube-assembly in a 'traditional' setup with basic wings attached, (b) our optimized two-wing design and (c) our optimized four-wing designs during pool testing. (d) An overview of the internal hardware assembly, including buoyancy engine and mass-shifter. (e) An illustration of how the outer shells assemble with the internal hardware.