Predictive Framework for the Optimization of Soft Bending Actuators

We present a machine learning-based framework leveraging artificial neural networks (ANNs) to optimize the mechanical performance of micro soft bending actuators while reducing analysis and experimentation time. A dataset of 10,000 randomly generated actuator designs was automatically created and analyzed using finite element analysis (FEA). The predictive model, employing a multilayer perceptron (MLP) regression algorithm trained with stochastic gradient descent (SGD), iteratively generated 3D-printable actuator designs until the expected bending angle matched the target. Fabricated designs were experimentally validated via 3D printing, achieving a low error of 1.5% compared to predicted bending angles under 30 kPa of applied pressure. These findings confirm that machine learning-based frameworks can significantly reduce soft actuator development time and effort.