Data-driven design of MXene aerogels with programmable mechanical performance via active learning and collaborative robots

Thumbnail Image

Publication or External Link





There is a solid demand for developing intelligent pressure sensing materials for the next generation of soft machines and robots. The piezoresistive pressure sensor requires a high sensitivity within a specific pressure range and possesses superior mechanical stability. Ti3C2Tx MXene-based aerogels with high electrical conductivities have been demonstrated as promising piezoresistive materials for the fabrication of intelligent pressure sensors for diverse sensing applications, from ultra-low stress vibration detection to irregular object grasping. MXene aerogels' piezoresistive behaviors can easily be tuned by changing the fabrication recipes that affect micro/nanostructures. Although many techniques have been reported for fabricating MXene aerogels for specific detection limits, the influence of the interplaying factors and their effect on the aerogels' structures and mechanical properties are not clearly understood. To achieve the custom design for pressure sensors for any given sensing windows and mechanical requirements, understanding the complex correlations between fabrication recipes, aerogel microstructures, and mechanical properties becomes necessary. Since traditional trial-and-error approaches require the production and manual processing of a large amount of data and, therefore, are highly time-consuming. Also, it is impossible to use a trial-and-error-based approach to study multi-dimensional design space as the one needed to construct an enormous amount of MXene-based aerogel sensors. Machine learning is a powerful and versatile tool that uses data-driven computation to uncover underlying trends and complex correlations. Machine learning requires a data-rich system to study the correlations and make accurate analyses and predictions. As the quality and size of the data obtained from the literature remain narrow and biased, it becomes essential to design high-throughput experiments to supply high-quality data to develop prediction models via machine learning. In this presentation, we adopt a hybrid strategy using wet-lab experiments, a machine learning framework, and collaborative robot assistance to build up a prediction model and uncover the underlying design principles to understand the mechanical properties of MXene-based aerogel sensors. Three functional materials (i.e., Ti3C2Tx MXene nanosheets, cellulose nanofibers, and gelatin), and one crosslinker (i.e., glutaraldehyde), are used for the fabrication of piezoresistive aerogels. First, a support-vector machine classifier is trained with 264 different compositions to confirm a feasible fabrication regime. Second, 160 piezoresistive aerogels with various recipes and morphologies are fabricated through active learning loops. Third, through data analyses, data-driven design principles for piezoresistive aerogels were uncovered and validated via in situ microscopic studies. Through this study, we make a crucial discovery about the roles of mass loading and cellulose nanofiber concentration on the mechanical properties of the resulting aerogels. Finally, we demonstrate how the implementation of collaborative robots can accelerate the prediction model construction.