Automatic MXene Strain Sensor Design via Active Learning and Data Argumentation for Soft Machines

Emerging soft machines require high-performance strain sensors to achieve closed-loop feedback control. Predicting the performance of a soft robotic sensor from its composition and morphology is nearly impossible with traditional computational approaches. Machine learning (ML) is a versatile tool to uncover complex correlations between fabrication recipes and sensor performance, yet the limited acquisition rate of high-quality data hinders the development of high-accuracy prediction models at the device level. In this talk, I will demonstrate our recent work of using an ML model to predict device-level performance and recommend new material compositions for soft machine applications. I will present a three-stage ML framework to construct a prediction model capable of automating the design of strain sensors across a wide strain range from <0.5% to 350%. First, a support-vector machine classifier is trained by using 351 compositions of various nanomaterials, including Ti3C2TX MXene nanosheets, single-walled carbon nanotubes, and polyvinyl alcohol. Second, through 12 active learning loops, 125 strain sensors are stagewise fabricated to enrich the multi-dimensional dataset. Third, data augmentation is implemented to synthesize >10,000 virtual data points followed by genetic algorithm-based selection to optimize the prediction accuracy of ML model. An ultimate prediction model is finally constructed and able to (1) predict sensor characteristics based on fabrication recipes and (2) recommend feasible fabrication recipes for adequate strain sensors. As final demonstrations, model-suggested strain sensors are integrated into soft gripper and batoid-like swimmer to endow them with real-time sensing capabilities.