Browsing by Subject "machine learning methods"

Sort by: Order: Results:

Now showing items 1-4 of 4
  • Dominguez-Gutierrez, F. J.; Byggmastar, J.; Nordlund, K.; Djurabekova, F.; von Toussaint, U. (2021)
    In this work, we study the damage in crystalline molybdenum material samples due to neutron bombardment in a primary knock-on atom (PKA) range of 0.5-10 keV at room temperature. We perform classical molecular dynamics (MD) simulations using a previously derived machine learning (ML) interatomic potential based on the Gaussian approximation potential (GAP) framework. We utilize a recently developed software workflow for fingerprinting and visualizing defects in damaged crystal structures to analyze the Mo samples with respect to the formation of point defects during and after a collision cascade. As a benchmark, we report results for the total number of Frenkel pairs (a self-interstitial atom and a single vacancy) formed and atom displacements as a function of the PKA energy. A comparison to results obtained using an embedded atom method (EAM) potential is presented to discuss the advantages and limits of the MD simulations utilizing ML-based potentials. The formation of Frenkel pairs follows a sublinear scaling law as xi ( b ) where b is a fitting parameter and xi = E (PKA)/E (0) with E (0) as a scaling factor. We found that the b = 0.54 for the GAP MD results and b = 0.667 for the EAM simulations. Although the average number of total defects is similar for both methods, the MD results show different atomic geometries for complex point defects, where the formation of crowdions by the GAP potential is closer to the DFT-based expectation. Finally, ion beam mixing results for GAP MD simulations are in a good agreement with experimental mixing efficiency data. This indicates that the modeling of atom relocation in cascades by machine learned potentials is suited to interpret the corresponding experimental findings.
  • Niittynen, Pekka; Heikkinen, Risto K.; Luoto, Miska (2020)
    Proceedings of the National Academy of Sciences of the United States of America 117: 35, 21480-21487
    The Arctic is one of the least human-impacted parts of the world, but, in turn, tundra biome is facing the most rapid climate change on Earth. These perturbations may cause major reshuffling of Arctic species compositions and functional trait profiles and diversity, thereby affecting ecosystem processes of the whole tundra region. Earlier research has detected important drivers of the change in plant functional traits under warming climate, but studies on one key factor, snow cover, are almost totally lacking. Here we integrate plot-scale vegetation data with detailed climate and snow information using machine learning methods to model the responsiveness of tundra communities to different scenarios of warming and snow cover duration. Our results show that decreasing snow cover, together with warming temperatures, can substantially modify biotic communities and their trait compositions, with future plant communities projected to be occupied by taller plants with larger leaves and faster resource acquisition strategies. As another finding, we show that, while the local functional diversity may increase, simultaneous biotic homogenization across tundra communities is likely to occur. The manifestation of climate warming on tundra vegetation is highly dependent on the evolution of snow conditions. Given this, realistic assessments of future ecosystem functioning require acknowledging the role of snow in tundra vegetation models.
  • Chumachenko, Kateryna; Männistö, Anssi; Iosifidis, Alexandros; Raitoharju, Jenni (IEEE, 2020)
    IEEE Access 8 (2020)
    In this paper, we demonstrate the benefits of using state-of-the-art machine learning methods in the analysis of historical photo archives. Specifically, we analyze prominent Finnish World War II photographers, who have captured high numbers of photographs in the publicly available Finnish Wartime Photograph Archive, which contains 160,000 photographs from Finnish Winter, Continuation, and Lapland Wars captures in 1939-1945. We were able to find some special characteristics for different photographers in terms of their typical photo content and framing (e.g., close-ups vs. overall shots, number of people). Furthermore, we managed to train a neural network that can successfully recognize the photographer from some of the photos, which shows that such photos are indeed characteristic for certain photographers. We further analyzed the similarities and differences between the photographers using the features extracted from the photographer classifier network. We make our annotations and analysis pipeline publicly available, in an effort to introduce this new research problem to the machine learning and computer vision communities and facilitate future research in historical and societal studies over the photo archives.
  • Ahola, Jaakko; Raatikainen, Tomi; Alper, Muzaffer Ege; Keskinen, Jukka-Pekka; Kokkola, Harri; Kukkurainen, Antti; Lipponen, Antti; Liu, Jia; Nordling, Kalle; Partanen, Antti-Ilari; Romakkaniemi, Sami; Räisänen, Petri; Tonttila, Juha; Korhonen, Hannele (Copernicus Publ., 2022)
    Atmospheric chemistry and physics
    The number of cloud droplets formed at the cloud base depends on both the properties of aerosol particles and the updraft velocity of an air parcel at the cloud base. As the spatial scale of updrafts is too small to be resolved in global atmospheric models, the updraft velocity is commonly parameterised based on the available turbulent kinetic energy. Here we present alternative methods through parameterising updraft velocity based on high-resolution large-eddy simulation (LES) runs in the case of marine stratocumulus clouds. First we use our simulations to assess the accuracy of a simple linear parameterisation where the updraft velocity depends only on cloud top radiative cooling. In addition, we present two different machine learning methods (Gaussian rocess emulation and random forest) that account for different boundary layer conditions and cloud properties. We conclude that both machine learning parameterisations reproduce the LES-based updraft velocities at about the same accuracy, while the simple approach employing radiative cooling only produces on average lower coefficient of determination and higher root mean square error values. Finally, we apply these machine learning methods to find the key parameters affecting cloud base updraft velocities.