Machine-learning-accelerated design of functional structural components in deep-sea soft robots

Abstract

To explore the deepest regions of the ocean with high flexibility and environmental adaptability, deep-sea soft robots have been developed recently. One prominent example is the self-powered soft robot that successfully operated in the Mariana Trench at a depth of 11,000 meters. Notably, many functional electronic components such as resistive elements, capacitors, and crystal oscillators may fail under extreme hydrostatic pressure, posing significant challenges for the practical massive deployment of deep-sea soft robots. Consequently, designing miniature pressure vessels on the printed circuit board to protect these vulnerable electronic components is vital for enhancing the reliability of deep-sea soft robots. However, traditional structure design methods – which often rely on theoretical analysis, experimental testing and numerical simulations – are often costly and time-consuming, especially for design problems in high-dimensional design spaces. Herein, we demonstrate a machine-learning-accelerated design method for devising miniature pressure vessels for vulnerable electronic components in deep-sea soft robots. Machine learning algorithms including decision trees and neural network models are employed and compared. The resulting design algorithm can predict whether a specific design can survive the deep-sea hydrostatic pressure with high accuracy in ∼0.35 ms, roughly seven orders of magnitude faster than traditional design methods.