Browsing by Author "Turpeinen, Aleksi"

Now showing 1 - 1 of 1

Deep Learning Model Training for 3D Molecules in Atomic Force Microscopy
(2024-02-07) Turpeinen, Aleksi
Perustieteiden korkeakoulu | Bachelor's thesis
Atomic Force Microscopy (AFM) is a nanoscale technique that offers the capability to capture high-resolution images of single atoms or molecules. This is done by bringing a small, metallic cantilever with a carbon monoxide (CO) molecule attached to its tip close to the examined surface in ultra-vacuum conditions close to absolute zero temperature. The cantilever is driven to oscillate with a specific frequency near the surface. The interaction forces between the atoms on the surface and the CO molecule at the cantilever's tip induce modifications to the oscillation frequency. These changes provide information about the surface's molecular and atomic structure. In this work, deep learning was used with neural networks in order to improve the resolution and clarity of simulated AFM images and gain more information about their 3D molecular structure. The accuracy of the neural network was measured with a loss function computed with the mean squared error method, which was minimized with gradient descent. Instead of real-life AFM images, a Probe Particle Model was used to simulate an AFM system using the Lennard-Jones potential and the Coulomb force. Two large datasets of simulated AFM images were given to a neural network to train it. One dataset was smaller than the other, but had a large amount of rotations for the molecules. The other dataset was larger and contained molecules with heavier elements such as bromine and chlorine. After this phase, separate simulated AFM images were fed to the trained neural network to test the model. The neural network training required significant computational resources, but using graphics processing units (GPU) on the Aalto University Triton server greatly sped up the training process. The neural network demonstrated a significant improvement in the simulated AFM images. This enhancement made individual atoms within the molecules distinctly visible, and the geometric configuration of the observed molecules easily ascertainable. By combining the precision of AFM with the computational power of neural networks, this work advances our understanding of molecular and atomic landscapes at the nanoscale.