Official websites use .gov
A .gov website belongs to an official government organization in the United States.
Secure .gov websites use HTTPS
A lock (
) or https:// means you’ve safely connected to the .gov website. Share sensitive information only on official, secure websites.
AI-STEM: A Revolutionary Approach to High-resolution Materials Characterization
Description
Overview
Semiconductor chips manufactures use imaging at the nanoscale to find defects that could affect chip performance. The Scanning Transmission Electron Microscope (STEM) is one of the most powerful tools for this work, but the high-energy electron beam used in STEM can damage or destroy chips, especially when used to take a large number of measurements. This project uses AI to dramatically reduce the number of measurements needed to produce accurate 3D reconstructed images of semiconductor devices and thus reduce the risk of chip damage.
Background
Manufacturing semiconductor chips requires imaging at the nanoscale to find defects that could affect chip performance. STEM is one of the most powerful tools for this work, it can reveal the atomic structure of materials and map their composition, chemistry, and physical properties. Combined with tomographic techniques, these measurements are performed in three dimensions.
However, the high-energy electron beam used in STEM can permanently damage or destroy the chip being inspected. Each measurement carries a cost, and collecting the hundreds of images that traditional 3D reconstruction methods require will harm the chip which changes the image measurement after each acquisition. The instruments used for these measurements are also expensive, making analysis time a premium resource.
This project uses AI to dramatically reduce the number of measurements needed to produce accurate 3D images of semiconductor devices.
Approach
Rather than collecting images at every possible angle, which would damage a chip, the team developed methods that can reconstruct a complete, high-quality 3D image from as few as 35 carefully chosen angles, instead of the 180 that conventional methods require.
The approach combines two AI techniques: one that removes noise from the limited measurements collected, and another that fills in the missing data by predicting what the uncollected measurements would have shown. To train these AI systems, the team generated thousands of synthetic chip images, since real ground truth data cannot be collected without damaging actual chips.
A key challenge was determining which 35 measurement angles, out of 180 possible, would capture the most useful information. The team tested several statistical methods to identify the most informative angles and found that a carefully optimized selection could preserve nearly all of the information contained in a full set of measurements.
The AI models were built using a neural network architecture well-suited to image processing tasks, and were trained on 10,000 synthetic variations of a transistor design. The networks work in two stages: first cleaning up noisy measurement data, then filling in the gaps left by the uncollected angles.
Impact
The team demonstrated that their AI-based approach can match the image quality of conventional reconstruction methods that use the full set of full 180 angle measurements, while using only 35. This represents a roughly 80% reduction in the number of measurements needed, significantly reducing the risk of damage to the chip during inspection and shortening analysis time.