Interfacing Machine Learning and Experimental Methods for Surface Structures (IMPRESS)

July 11, 2018 to July 13, 2018
Location : CECAM-AT


  • Oliver T. Hofmann (Institute of Solid State Physics, Graz University of Technology, Austria)
  • Milica Todorovic (Department of Applied Physics, Aalto University, Finland)
  • Patrick Rinke (Aalto University, Helsinki, Finland)






The structure of a surface determines its properties and its function. This holds especially true for organic overlayers and organic films grown on inorganic surfaces. The adsorption of molecules on surfaces is a complex processes determined by multiple physical and chemical phenomena, such as multipole-multipole interactions, van-der-Waals interactions, covalent-bond formation and charge-transfer. Thus, one and the same system often exhibits various different adsorption motifs, depending on the details of the deposition conditions [1] Such polymorphs exhibit different electronic properties, establishing a growing interest in elucidating the details of the adsorption structure.[2] Conversely, understanding the origin as well as the impact of structural polymorphism at surfaces and interfaces will facilitate a rational design and engineering of particular motifs to tailor structure-dependent functionalities.[3]
Determining or predicting the structure of organic ensembles on surfaces is a challenging problem, both experimentally and theoretically. Experimental methods often lack the resolution to resolve necessary detail in the molecular structure or arrangement or are too slow to scan larger arrangements. Quantum theories can resolve the structure of individual molecules, but the computational costs of addressing molecular ensembles remain prohibitive. (Semi-)classical theories can bridge the scales to polymorphism and morphological questions, but lack the quantum mechanical component that is often required to accurately describe the many competing interactions.
In the past years, computational structure search methods inorganic/organic interfaces have made tremendous progress. This progress can be mostly attributed to the advent of machine learning methods in material science. Machine learning is now used to obtain cheaper, more accurate total energies (e.g. by machine-learned force fields [4]), accelerates established techniques (such as genetic algorithms via clustering [5] or cluster expansion via compressed sensing [6]), and even to develop completely new search strategies (e.g. via Bayesian learning [7]). Despite the stunning success of computational structure search, most studies still focus on small model systems. A routine application to interfaces that are directly relevant to engineering problems is still challenging because of the occurrence of “real-world problems”, i.e. the imperfections of the interfaces. These could be, for example, stacking faults, dislocations, substrate step-edges, molecular impurities, misaligned molecules, grain boundaries and rotational domains, dislocations, or many more.
Experimental groups also employ algorithms to convert their measured data into physical interpretations, i.e. surface structures [8]. Naturally, these algorithms have evolved to be very robust and highly tolerant for defects. Still, they are specialized towards specific methods, and often (because a single experimental method almost always only contains incomplete information about the system measured), do not provide unique, unambiguous solutions. Rather, the results should be compared against computationally determined structures for plausibility and correctness. Moreover, the analysis of experimental data is often controversial and contains assumptions which themselves depend on the observed surface structure [9], and thus needs to be performed iteratively. As a result, the experimental algorithms are often just as sophisticated as computational structure search techniques. Since computational and experimental algorithms have, however, evolved to solve complimentary challenges efficiently, the respective method developers could greatly benefit from each other’s knowledge and expertise.

The proposed workshop addresses the growing interest in interface morphologies. Experimental characterization and theoretical simulation techniques have progressed rapidly and it is now time to channel the novel developments from the different surface science sub -disciplines into dedicated workshop. The proposed composition of invited speakers combines experimentalists with theoreticians to promote synergetic solutions for current and future interfaces.

In particular, we will focus on the following questions, some of which were raised by the participants during the previous IMPRESS workshop

> Defects and grain boundaries: Which kind of defects need to be accounted for in computational studies? To what extend are defects and their energies relevant for the observed phase? Are defects a frequent mechanism for organic overlayers to reduce strain induced by the substrate lattice?

> Commensurability: Band-structure calculations of interfaces always assume fully commensurate interfaces. Many interfaces, and in particular technologically relevant ones, show point-on-line or point-on-point commensurability, or are fully incommensurate. How can such interfaces be considered within first-principle calculations, without further increasing the already almost intractably large search space?

> Thermodynamics versus kinetics: Monte-Carlo studies notwithstanding, most computational efforts focus on the thermodynamically most stable structure. How can we make sure that an experimentally observed structure is indeed thermodynamically most stable, and not metastable? Is it possible to define clear target properties that allow computational structure search methods to be geared towards metastable structures?

> Processing conditions: In experiment, the obtained polymorph often depends sensitively on the processing conditions, in particular when solution-based techniques, such as bar meniscus shearing or doctor-blading, is employed. (How) should structure search techniques account for the processing? Since these structures are often not in thermodynamic equilibrium, what is the property that should be optimized? (See previous question)

> Error bars: The analysis of experimental always leaves some room for interpretation, at least in the form of error bars that are inevitably associated with the data. How should structure search techniques account for such error bars?

> Multilayer and morphology: While efficient algorithms exist to predict the structure of bulk materials or for interfaces (i.e., monolayers), strategies to determine the geometry of multilayers or thin films are scarce. What is a good strategy to consider the second, third, etc. layer at the interface? Is it plausible to only consider the first monolayer and neglect the substrate? How do we deal with “cannibalizing” structures, where the first (wetting) layer disappears after deposition of additional material?

> Polymorph identification and retrieval: There is presently no established method to label and uniquely identify a given surface structure. This poses a significant challenge for the comparison between theory and experiment and makes a retrieval of already proposed structure from literature or material databases, such as NOMOAD, almost impossible. Is it possible to develop a sensible, easy to apply labelling scheme to interfaces?



[1]: Llinàs, A. & Goodman, J. M. Polymorph control: past, present and future. Drug Discov. Today 13, 198–210 (2008).
[2]: Jones, A. O. F., Chattopadhyay, B., Geerts, Y. H. & Resel, R. Substrate-Induced and Thin-Film Phases: Polymorphism of Organic Materials on Surfaces. Adv. Funct. Mater. 26, 2233–2255 (2016).
[3]: Chung, H. & Diao, Y. Polymorphism as an emerging design strategy for high performance organic electronics. J Mater Chem C 4, 3915–3933 (2016).
[4]: Botu V.,Batra R., Chapman J., & Ramprasad R. Machine Learning Force Fields: Construction, Validation, and Outlook, J. Phys. Chem. C, 121 (1) pp 511–522 (2017)
[5]: Zhang. J, Chung. H & Lo. W. L, "Clustering-Based Adaptive Crossover and Mutation Probabilities for Genetic Algorithms", IEEE Transactions on Evolutionary Computation 11 (3), 326–335 (2007).
[6]: Nelson L.J., Ozolins V., Reese S., Zhou F., Hart G.L.W. Cluster expansion made easy with Bayesian compressive sensing, Phys. Rev. B 88, 155105 (2013)
[7]: Packwood D.M., Han, P., Hitosugi T., Chemical and entropic control on the molecular self-assembly process, Nature Communications 8, 14463 (2017)
[8]: Viana M.L., dos Reis D. D., Soares E.A., Van Hove M.A., Moritz W & de Carvalho V. E., Novel genetic algorithm search procedure for LEED surface structure determination J. Phys.: Condens. Matter 26 225005 (2014)
[9]: Zegenhagen J. & Kazimirov A. (editors), The X-Ray Standing Wave Technique, Principles and Applications, Series on Synchrotron Radiation Techniques and Applications: Volume 7