Multi-View Neural 3D Reconstruction of Micro-/Nanostructures with Atomic Force Microscopy

TL;DR

This work tackles the inability of conventional AFM to recover full 3D micro-/nanostructures due to tip-sample convolution and limited vertical information. MVN-AFM combines multi-view AFM scanning, an EM-like iterative data-alignment and artifact-masking stage, and neural implicit surface reconstruction by learning a Signed Distance Field $s(x; \theta)$ via differentiable volume rendering. The SDF-based model is trained with a depth supervision loss $L = L_{depth} + \lambda L_{reg}$, where $L_{depth} = \frac{1}{b} \sum_p (\hat{d}_p - d_p)^2$ and $L_{reg} = \frac{1}{bm} \sum_{p,q} (||\mathbf{n}_{pq}|| - 1)^2$, enabling artifact-free surface extraction through Marching Cubes. Across two experimental demonstrations (Two-photon Lithography microstructures and nanoparticles) and simulated data, MVN-AFM achieves accurate 3D reconstructions with substantially reduced topography error and improved pose accuracy, highlighting its cost-effectiveness and generalizability.

Abstract

Atomic Force Microscopy (AFM) is a widely employed tool for micro-/nanoscale topographic imaging. However, conventional AFM scanning struggles to reconstruct complex 3D micro-/nanostructures precisely due to limitations such as incomplete sample topography capturing and tip-sample convolution artifacts. Here, we propose a multi-view neural-network-based framework with AFM (MVN-AFM), which accurately reconstructs surface models of intricate micro-/nanostructures. Unlike previous works, MVN-AFM does not depend on any specially shaped probes or costly modifications to the AFM system. To achieve this, MVN-AFM uniquely employs an iterative method to align multi-view data and eliminate AFM artifacts simultaneously. Furthermore, we pioneer the application of neural implicit surface reconstruction in nanotechnology and achieve markedly improved results. Extensive experiments show that MVN-AFM effectively eliminates artifacts present in raw AFM images and reconstructs various micro-/nanostructures including complex geometrical microstructures printed via Two-photon Lithography and nanoparticles such as PMMA nanospheres and ZIF-67 nanocrystals. This work presents a cost-effective tool for micro-/nanoscale 3D analysis.

Figures (16)

• Figure 1: The pipeline of MVN-AFM.a First, we place various micro-/nanostructures on a rotatable tilt stage. Second, we rotate the turntable and measure the vertical heights by a conventional AFM, resulting in a set of multi-view AFM images with many artifacts. b Input raw AFM images with artifacts and iterate two sub-steps. In the data alignment process, data judged as artifacts are eliminated before alignment, and the poses of multi-view images are updated. In the mask-solving process, the solved poses transform the multi-view data, and the data consistency is cross-validated to solve the mask of artifacts. c The posed and masked multi-view AFM images are used to train a neural network representing a signed distance field in space by the differentiable volume rendering technique (Supplementary \ref{['fig: network_pipeline']}). d The 3D surface model extracted from the signed distance field, and corresponding topography images without artifacts.
• Figure 1: The limitations of conventional AFM scanning.a Illustration of conventional AFM scanning process. AFM obtains topography information in the vertical direction by the interaction between the probe and the sample. However, when the side of the probe, rather than the probe's tip, touches the sample, AFM cannot get accurate height information. This phenomenon, which leads to artifacts in the AFM images, is called tip-sample convolution. b The 3D model of the conventional AFM scanning result. The scanning result is the combination of an incomplete structure surface (the blue part) and artifacts (the red part).
• Figure 2: MVN-AFM reconstructs the surface models of two-photon lithography microstructures.a-f SEM photos of TPL microstructures. g-l 3D models of TPL microstructures' conventional AFM scanning data. m-r 3D models of TPL microstructures reconstructed by MVN-AFM. g, h, m, n Include the cross-section profiles in the x-z plane. i-l, o-r Include the cross-section profiles in the x-y plane. More visualizations can be found in Supplementary \ref{['fig: TPL_SEM_top_down']} and Supplementary Movie 1.
• Figure 2: The tilt stage for multi-view AFM scanning.a, b The design model of the tilt stage and the turntable in its center. c The photos of the tilt stage and the AFM multi-view scanning process.
• Figure 3: MVN-AFM reconstructs the surface models of nanoparticles.a, b SEM photos of nanoparticles in the overhead view. c-f SEM photos of nanoparticles in the tilt view. g-j 3D models of nanoparticles' conventional AFM scanning data. k-n 3D models of nanoparticles reconstructed by MVN-AFM. i, m The local zoom reveals two stacked-up nanocrystals. More visualizations can be found in Supplementary Movie 2.