users bring cool science to the beamline

experimental data is reduced

data analysis is performed

nature paper is written and published

what can we automate?

  • Users bringing the cool science: 🤦
  • Data reduction: ✅
  • Data analysis: 🤞
  • Writing papers: 😔
  • x-ray reflectometry reduction is well studied

    Previous literature focuses on point detectors, while modern instruments utilise area detectors
    A. Gibaud, G. Vignaud, & S. K. Sinha. Acta Cryst., A49, 642-648, 1993
    F. Salah, B. Harzallah, & A. van der Lee. J. Appl. Crystallogr., 40, 813-819, 2007.

    not re-inventing the wheel, but re-building it

    this will enable automated reduction

  • Only a single processing pipeline for a given experiment
  • The processing is modular
  • Coming to i07 in October 2019
  • reflectometry analysis is model-dependent

    Can we automate the model selection?

    machine learning for model selection

    ml-driven classification

    ml-driven classification

    1. A. R. McCluskey, et al.. Phys. Chem. Chem. Phys., 21, 6133-6141, 2019.
    2. Y. Gerelli. Phys. Rev. Lett., 122, 248101, 2019.
    3. V. Rondelli, G. Fragneto, S. Motta, E. Del Favero, L. Cantù, J. Phys. Conf. Ser., 340, 012083, 2012.

    convolutional neural network

    Very quick and accurate for "model" data;
    2 epochs → 100 % validation accuracy

    how robust is our model?

    Monolayer Bilayer Floating Bilayer
    DPPC Validation 100 % 100 % 100 %
    DXPC Validation 100 % 100 % 100 %

    with ml we can advise analysis

    We can create generic analysis interfaces which the machine learned classification would suggest

    jupyter notebook

    This could be in the form of a Jupyter Notebook that you could run from your office

    automated script

    Or an automated analysis script could be run on well tagged data

    Penny and Sadie, good dogs