Machine learning for the identification of molecular conformationsWe can overcome the information gap by acquiring large amounts of data while manipulating the molecule a small region around its current conformation. As long as this manipulation is fully reversible, the conformation is essentially not changed. We will implement machine learning solutions (such as (un)supervised learning, kernel ridge regression, support vector machines etc.) for the various data analysis tasks required in this context. Supervised machine learning which, based on wavelet analysis, detects discontinuities in the experimental signals that indicate (irreversible) conformational relaxations. Unsupervised machine learning to structure small segments of experimental data into larger units. Acquiring large amounts of simulation data using a reliable and accurate molecular mechanics model. Supervised machine learning (support vector machine) to determine the shape of the region within which manipulation is reversible for experimental and simulated data. Kernel regression to interpolate measured and simulated data in each region. An algorithm to assign experimental regions to simulated ones on a bipartite graph via measurements of manifold distances. This creates a map of all possible manipulation processes.
White Paper: Understanding Many-Particle Systems with Machine Learning, Inst. of Pure and Appl. Mathematics (IPAM) (2017)
▶February 15th, 2019We welcome Joshua Scheidt, computer science student at Maastricht University, who joins the MoMaLab to do his master thesis.
▶February 4th, 2019Hand-controlled STM-based atomic manipulation with real-time visual feedback from an MD simulation Read at Beilstein..
▶October 25th, 2018We have been granted 2 Mio core‑h on JURECA for our simulation work on molecular adsorption and manipulation.
▶September 6th, 2018We are happy that MSc Michael Maiworm joins us till January to work on molecular control algorithms.