Deep-Learning Inversion Maps Arbitrary Design Images to Low-Cost, Efficient Nanofabrication.

Abstract

Rapid, low-cost production of user-defined nanoscale patterns is vital for prototyping in energy, biomedical, and information technologies. Yet top-down lithography is prohibitively expensive, and bottom-up self-assembly affords limited design freedom. Shadow sphere lithography (SSL) enables wafer-scale nanopatterning using inexpensive microsphere masks but typically relies on labor-intensive trial-and-error to translate desired structures into fabrication parameters. This task is reformulated as a fabrication-oriented image-to-recipe translation solved with deep learning. Analytic shadow projection equations generated more than 4.5 × 10⁶ synthetic patterns and parameter pairs to train a bidirectional convolutional block attention network. By jointly extracting deep features and a redesigned loss function, the network learns the mapping between target nanostructures and viable processing conditions. For unseen designs, the model returns a full fabrication recipe in under one second on a consumer GPU, achieving 91% parameter accuracy and a Pearson correlation of 0.95 ± 0.01 between the target and predicted structures. Deployed as a Web application, the framework successfully guides the fabrication of diverse nanoscale patterns within the SSL design space. Transforming a weeks-long, high-cost workflow into an automated one-click operation, the framework enables fast, on-demand, and cost-efficient nanomanufacturing.