Institute of Solid State Physics


SS23WS23SS24WS24SS25      Guidelines for Master Students

AMAN-SPM: Autonomous Molecular and Atomic Nanofabrication via SPM
Bernhard Ramsauer
Institute of Solid State Physics
https://tugraz.webex.com/tugraz/j.php?MTID=md822433cba4a614777b87ed063271b28
11:15 - 12:15 Wednesday 27 November 2024 

In the AMAN-SPM project, I will autonomously synthesize molecules on metal surfaces and arrange them into functional nanostructures that are inaccessible via conventional chemistry by developing an innovative, integrated machine-learning-based approach. The new molecules will be obtained by synthesizing reactive precursor species, precisely positioning and correctly orienting them in close proximity to each other, and inducing bond formation via voltage pulses applied with the tip of a scanning probe microscope (SPM).

The biggest challenge in achieving this goal is the complex nature of the molecule-surface and molecule-tip interaction processes. To perfectly control on-surface synthesis, it is necessary to understand the outcome of each manipulation. However, predicting the outcome is far from trivial as the interaction processes strongly depend on the manipulation parameters (i.e., relative position of the tip and voltage applied) and are non-deterministic. Furthermore, the underlying physics is typically unknown and often extremely difficult to understand. Thus, learning how to manipulate molecules manually is tedious, time-consuming, and often outright impossible for a human expert.

The AMAN-SPM project will be a breakthrough in performing on-surface synthesis. The SPM will operate based on decisions made by a machine learning algorithm and, hence, act autonomously. The ideal machine learning framework to learn on-surface synthesis is reinforcement learning (RL). A priori, it does not require prior knowledge of a physical model of the interaction processes, as it learns the optimal manipulation parameters for the individual synthesis steps directly by controlling the SPM and conducting experiments while meticulously analyzing the outcomes. The method is thus easily transferable to different systems and can readily be scaled up from the synthesis of single molecules to the construction of more complex nanostructures. Additionally, since the learning procedure inherently requires thousands of precisely analyzed data points for every performed manipulation, we can combine the analysis with ab-initio calculations to obtain insight into the interaction processes.