Near-perfect replication on amorphous alloys through active force modulation based on machine learning/neural network parameter prediction

Abstract

As a microforming technique, micro/nano-structural replication possesses advantages of high precision and efficiency. With the remarkable superplasticity in the supercooled liquid region, amorphous alloys or metallic glasses (MGs) are regarded as ideal materials for miniature fabrication. However, due to the intrinsic metastable nature of supercooled liquids, the design of imprinting processes for MGs poses a challenge. In the past, process parameters have largely relied on trial-and-error strategies. In this work, a low-frequency active force modulation system is employed to apply a stable, precise, and real-time feedback stress field for imprinting of MG samples. Low-frequency vibrations can facilitate the filling of microstructures on the template surface by reducing the effective viscosity of the supercooled liquid. With the dataset composed of orthogonal experiments, a machine learning strategy based on back-propagation (BP) neural networks was utilized to construct a 3D parameter space for temperature, stress, and time, and to predict the corresponding filling ratio. Furthermore, the optimal combination of imprinting process parameters was identified, and its filling ratio was experimentally validated to reach as high as 0.94. The near-perfect replication of microstructures confirms the superiority of the active force modulation system and the data-driven strategy of machine learning-assisted parameter design. At the same time, this one-step microforming process provides a new approach to addressing the accuracy-cost trade-off dilemma in precision manufacturing.