(403g) Artificial Intelligence-Assisted Full-Map Understanding of Strain Sensing Devices and Designated Predictions for Soft Robotic Systems

Compliant strain sensing devices are widely applied in real-time mechanical feedback for soft robotic systems. However, it remains challenging to reach customized sensing devices with high sensitivity coordinated with designated working windows due to the complex and unclear structure-composition-property relationships which are hard to predict and pre-design. In this work, through active learning and developed algorithms, we propose machine assisted full-map structure-composition-property understandings of strain sensors consist of classical nanocomponents of Ti₃C₂T_x MXene, carbon nanotube, and poly(vinyl alcohol). Important variables of nanocomponent weight loading, nanocoating thickness, and nanocoating surface morphology are systematically investigated, and an informative data base (5000+ experimental resistance points, 500+ nanocoatings, and 200+ sensing devices) is constructed for training of machine learning models. Besides, comprehensive algorithm optimizations are performed for better machine performance. Ultimately, the machine learning models show low average scattered numerical error of 24% and give customized predictions in 1) sensing performance in a given material recipe; 2) suitable material recipes of sensing devices with high sensitivity and designated working windows in soft robotic systems. The reported strategies here open new perspectives towards design-free high-performance strain sensing devices and customized applications in soft robotic systems.

More specifically, in our material systems, at least 33,700 non-repeating recipes are available for strain sensing devices fabrication although we limit the step length in each variables of material recipes. With such a complex system and very limited understanding, there are no available in silico simulation methods to interpret or predict the sensing devices. In addition, it is nearly impossible to screen the whole design space via Edison approach because the sensing needs are customized every time. Herein, we have proposed a data-driven framework to exploratively learn the full-map of the strain sensing devices. The data is collected with thirteen active learning rounds which investigate both the mean minimal distance (L₂)between suggestion points and existing data points, and the numerical uncertainty of well-trained models during each round. After a representative dataset is constructed, optimization strategies including GA combination, customized loss function, and data augmentation are used for improving the model performance. This best model has achieved a low average scattered numerical error ((Predicted value - Real value)/Real maximum value in device) of 24%.