Atomic pressure microscopy, or AFM, is a broadly used method that may quantitatively map materials surfaces in three dimensions, however its accuracy is proscribed by the dimensions of the microscope’s probe. A brand new AI method overcomes this limitation and permits microscopes to resolve materials options smaller than the probe’s tip.
“Correct floor peak profiles are essential to nanoelectronics improvement in addition to scientific research of fabric and organic methods, and AFM is a key method that may measure profiles noninvasively,” mentioned Yingjie Zhang, a U. of I. supplies science & engineering professor and the venture lead. “We’ve demonstrated learn how to be much more exact and see issues which might be even smaller, and we’ve proven how AI could be leveraged to beat a seemingly insurmountable limitation.”
Typically, microscopy strategies can solely present two-dimensional photos, basically offering researchers with aerial pictures of fabric surfaces. AFM supplies full topographical maps precisely displaying the peak profiles of the floor options. These three-dimensional photos are obtained by shifting a probe throughout the fabric’s floor and measuring its vertical deflection.
If floor options method the dimensions of the probe’s tip—about 10 nanometers—then they can’t be resolved by the microscope as a result of the probe turns into too giant to “really feel out” the options. Microscopists have been conscious of this limitation for many years, however the U. of I. researchers are the primary to offer a deterministic resolution.
“We turned to AI and deep studying as a result of we wished to get the peak profile—the precise roughness—with out the inherent limitations of extra typical mathematical strategies,” mentioned Lalith Bonagiri, a graduate scholar in Zhang’s group and the examine’s lead writer.
The researchers developed a deep studying algorithm with an encoder-decoder framework. It first “encodes” uncooked AFM photos by decomposing them into summary options. After the function illustration is manipulated to take away the undesired results, it’s then “decoded” again right into a recognizable picture.
To coach the algorithm, the researchers generated synthetic photos of three-dimensional constructions and simulated their AFM readouts. The algorithm was then constructed to remodel the simulated AFM photos with probe-size results and extract the underlying options.
“We really needed to do one thing nonstandard to attain this,” Bonagiri mentioned. “Step one of typical AI picture processing is to rescale the brightness and distinction of the pictures in opposition to some customary to simplify comparisons. In our case, although, absolutely the brightness and distinction is the half that’s significant, so we needed to forgo that first step. That made the issue rather more difficult.”
To check their algorithm, the researchers synthesized gold and palladium nanoparticles with identified dimensions on a silicon host. The algorithm efficiently eliminated the probe tip results and appropriately recognized the three-dimensional options of the nanoparticles.
“We’ve given a proof-of-concept and proven learn how to use AI to considerably enhance AFM photos, however this work is just the start,” Zhang mentioned. “As with all AI algorithms, we will enhance it by coaching it on extra and higher knowledge, however the path ahead is obvious.”
Extra data: Lalith Krishna Samanth Bonagiri et al, Exact Floor Profiling on the Nanoscale Enabled by Deep Studying, Nano Letters (2024). DOI: 10.1021/acs.nanolett.3c04712