Application of machine learning in molecular nanorobotics to obtain the required level of speed and accuracy in the computation of molecular manipulation experiments.
Could you introduce yourself, giving your affiliation, area of work, and of course, the project title?
The project ‘MOMONANO: A machine-learned prediction model for molecular nanorobotics’ is coordinated by Christian Wagner, group leader at the Peter Grünberg Institute (PGI-3) at Forschungszentrum Jülich, whose core expertise is in scanning probe microscopy (SPM) and the investigation of organic molecules. Together with his colleagues, he has developed very precise methods of single-molecule manipulation which, for example, allow individual molecules to become electric field sensors. He has the vision that the SPM could eventually be turned into an atomically precise molecular assembly machine, if combined in a sophisticated way with simulations.
Kristof T. Schütt is a senior researcher at the Berlin Institute for the Foundations of Learning and Data (BIFOLD). His research interests include interpretable neural networks, representation learning, generative models, and machine learning applications in quantum chemistry. Recently Schütt and his colleagues have developed SchNet, a neural network utilising continuous-filter convolutional layers, to predict chemical properties and potential energy surfaces of organic molecules.
In their collaboration, Schütt and Wagner showed in 2020 that autonomous nanorobotics is possible for a certain class of tasks when using reinforcement learning with sparse feedback.
In simple words, what specifically is your project about? And, how and why do you think it is a high risk, high gain endeavour?
The idea to freely control the atomic-scale structure of matter has intrigued scientists for many decades. The low temperature scanning probe microscope (SPM) has become the instrument of choice for this task since it allows the rearrangement of atoms and molecules on a surface. SPM-based nanofabrication in which molecules are ‘grabbed’ by the SPM tip and handled like LEGO building blocks is, however, impeded as the molecules cannot be imaged during manipulation and their conformation is therefore unknown. A solution to this problem can be achieved by comparing measured and calculated values of the force acting on the SPM tip that plays the role of the robotic actuator. If this is done continuously throughout the manipulation process it could enable inferring the actual molecular conformation by solving an inverse problem. Obviously, this approach requires the rapid availability of simulation data during the manipulation process and only sophisticated, computationally expensive atomistic calculations can provide the required accuracy. Currently, there exists no solution that combines the necessary levels of speed and accuracy in the computation of molecular manipulation experiments. The goal of MomoNano is therefore the creation of a machine-learned prediction model for molecular nanorobotics. To achieve this, we combine the expertise of experimental and theoretical physicists and machine learning experts. The main challenge of the project is the required combination of very high accuracy of the data predicted by the ML model and the large size of the system including a surface, a molecule, and an SPM tip. While sufficiently accurate ML models exist for individual gas-phase molecules, our project stretches their application to systems of much higher complexity. If we are successful, on the other hand, we will have created an essential component of a molecular manipulation setup which could eventually allow fabrication of functional molecular structures in a LEGO-type approach.
How important has the Helmholtz AI funding and platform been to carry out this project?
The funding received from the Helmholtz AI platform was absolutely essential to implement this project. Of particular importance is the collaborative nature of the funding scheme which allowed us to combine the expertise of three different research groups at two Helmholtz institutes and one university. This combination of experimental and theoretical physics, as well as machine learning is the key to our success. In that respect, the Helmholtz AI funding scheme which is inherently interdisciplinary, has a high risk - high gain profile and the short time for decisions about funding is perfect to execute projects of medium size but high urgency in the strategic framework of Helmholtz program-oriented funding.
Figure: Our project goal is to turn a scanning probe microscope into a precise assembly machine for individual molecules on a surface.