New Publication | Deep Learning Enables One-Click Nanofabrication Design

We are excited to share our latest publication in ACS Nano, titled “Deep-Learning Inversion Maps Arbitrary Design Images to Low-Cost, Efficient Nanofabrication.” 

This work was completed through an international collaboration between the University of Georgia and several leading research groups in China, including collaborators from Chongqing University, Zhejiang University, and the Eastern Institute of Technology, Ningbo. This collaboration brought together expertise in nanofabrication, shadow sphere lithography, deep learning, materials characterization, and inverse design.

Micro- and nanoscale patterns are essential for applications in photonics, sensing, biomedical devices, information technology, and advanced manufacturing. However, fabricating user-defined nanoscale structures remains difficult. Conventional top-down lithography provides high precision but requires expensive cleanroom tools, while bottom-up self-assembly is low-cost but offers limited design freedom.

In this work, we demonstrate a deep-learning-assisted inverse-design framework that converts arbitrary target images into practical fabrication recipes for shadow sphere lithography (SSL), a scalable and low-cost nanofabrication method based on colloidal sphere masks.

What We Did

We reformulated SSL design as an image-to-recipe translation problem. Instead of relying on manual trial-and-error to select sphere size, deposition angle, azimuthal direction, and deposition sequence, the model learns how to directly map a desired nanoscale pattern to a set of feasible fabrication parameters.

Using analytical shadow-projection equations, the team generated a large synthetic design space containing millions of pattern–parameter combinations. After removing duplicate and physically unrealistic cases, a curated dataset of 126,160 unique SSL structures was used to train a bidirectional convolutional block attention network, or Bi-CBAM.

The model includes both an inverse network and a forward network. The inverse network predicts fabrication parameters from a target structure, while the forward network reconstructs the resulting pattern from those parameters. This bidirectional design helps address the one-to-many nature of nanofabrication, where different fabrication recipes can sometimes produce very similar structures.

Key Results

The Bi-CBAM model successfully predicted SSL fabrication recipes for previously unseen patterns, achieving strong agreement between target and predicted structures.

The model reached a Pearson correlation coefficient of 0.95 ± 0.01 and a structural similarity index of 0.79 ± 0.15 on test patterns. It also predicted full fabrication recipes in less than one second on a consumer GPU.

To validate the approach experimentally, the predicted recipes were used to fabricate diverse nanoscale patterns, including geometric motifs, icons, and more complex free-form structures. The fabricated structures, characterized by AFM and SEM, preserved the major contours and nanoscale features of the predicted designs.

The framework was also deployed as a Web-based design tool, allowing users to upload an image and obtain SSL fabrication parameters and simulated nanoscale patterns within only a few seconds.

Why This Matters

This work transforms shadow sphere lithography from a largely trial-and-error process into a programmable, AI-assisted nanofabrication workflow. By connecting digital design images directly to low-cost fabrication recipes, this approach makes nanoscale pattern generation faster, more accessible, and more flexible. It provides a practical pathway for rapid prototyping of nanostructures for applications in metasurfaces, polarization optics, photonic devices, sensing platforms, and flexible nanomanufacturing. More broadly, this study demonstrates how deep learning can bridge the gap between design intent and physical fabrication, enabling a new generation of intelligent, low-cost, and on-demand nanomanufacturing tools.

Read the full article:Deep-Learning Inversion Maps Arbitrary Design Images to Low-Cost, Efficient Nanofabrication, ACS Nano 2026, 20, 14910–14922.DOI: 10.1021/acsnano.6c06350