Congrats, Noah! He won the freshly established Lieselotte Templeton Award for his Master thesis with the topic Improving Multidimensional Data Reduction and Analysis Algorithms for POWTEX: Raw-data Correction with Pseudo-Voigt-back-to-back-exponentials, Event Correlation Approaches and Machine Learning Supported Rietveld Refinement. Noah already started working in the POWTEX group of Prof. Richard Dronskowski and simulated neutron guides even before his Bachelor thesis. He completed his Bachelor’s degree, a research internship, and his master’s degree in the POWTEX group dealing with various subprojects [1]. Additionally, he gained experience on practical diffraction in the group of Prof. Ulli Englert [2, 3]. During his Master’s thesis and his PhD work, machine learning has become a key topic of his interest.
Please describe what you did during your Master thesis?
It is difficult to summarise everything in a few words – different topics, many experimental projects, but all more or less related to the future high-intensity neutron powder diffractometer POWTEX.
It can be broken down into three main topics: Data reduction, event correlation and machine learning. I was involved in optimising the POWTEX-specific data reduction procedure and implemented a new profile shape function to properly account for the asymmetry of the neutron source. During this work, we came up with the idea for a new experimental approach to further exploit all the detector’s capabilities – event correlation. It uses the unique four dimensionality (x, y, z and time) of the detector to reconstruct partial neutron trajectories and identify neutron events not originating from the sample. I have developed first preliminary results of this method as a proof of concept.
A key goal of implementing machine learning into our workflow is to lower the barrier to entry for new users into diffraction analysis with the increasingly complex – and information-rich – datasets generated by next-generation instruments such as POWTEX. In my thesis I designed an interface between a machine learning algorithm and the diffraction analysis software GSAS II.
I guess you programmed a lot. Who or what was your rubber duck?
Maybe not a proper rubber duck, but I really enjoy talking to my father (Prof. Klaus Nachtigall) about my scientific progress. He has (by his own admission!) little understanding of chemistry and is not as invested with coding. Thus, when I must present my work without going into every technical detail, it helps me find inconsistencies, similar to Rubber Ducking.
What was your biggest breakthrough during your thesis?
My biggest breakthrough was probably less a specific scientific achievement and more the personal progress I made during the master’s thesis. I went into the master’s thesis with a very rudimentary understanding of machine learning and can now build on my newly acquired knowledge for future scientific work. In this respect, the master’s thesis itself is the biggest “breakthrough”.
Did you have courses on Machine Learning during your studies?
I followed the DeepMind x UCL RL Lecture Series on Youtube by Hado van Hasselt on Reinforcement Learning. I also familiarised myself with A practical introduction to GNNs by Daniele Grattarola on Graph Neural Networks. Also worth mentioning is the workshop Do Research Like a Munchkin, which I attended. Although it is not strictly about machine learning, there is a close connection to the topics of Clean Coding and Agile Project Management.
Literature
- A. Houben, P. Jacobs, Y. Meinerzhagen, N. Nachtigall, R. Dronskowski: POWTEX visits POWGEN, 10.48550/arXiv.2110.12767
- S. van Terwingen, N. Nachtigall, B. Ebel, U. Englert: N-Donor-Functionalized Acetylacetones for Heterobimetallic Coordination Polymers, the Next Episode: Trimethylpyrazoles, Cryst. Growth Des. 21(5) 2962 (2021) 10.1021/acs.cgd.1c00122
- S. van Terwingen, N. Nachtigall, U. Englert: Synthesis and coordination to the coinage metals of a trimethylpyrazolyl substituted 3-arylacetylacetone, Z. Kristallogr. 10.1515/zkri-2021-2059