Probabilistic deep learning to automate the interpretation of multi-phase diffraction spectra

Physics-informed data augmentation was used to simulate realistic diffraction spectra, which formed the training set for a convolutional neural network that gives probabilistic predictions of compounds in experimental samples.

Scientific Achievement

We designed a deep learning algorithm to automate the identification of inorganic materials from experimentally measured X-ray diffraction spectra.

 Significance and Impact

The methods presented here may be integrated in experimental workflows to facilitate the high-throughput and autonomous discovery of inorganic materials.

Research Details

  • Simulated spectra were augmented with possible experimental artifacts left over from sample preparation and synthesis.
  • Off-stoichiometry was accounted for by simulating hypothetical solid solutions between known experimental phases.
  • To handle multi-phase mixtures, a branching algorithm was developed to identify the set of compounds that maximize the average probability of predictions given by the deep learning model.

Download this highlight 

DOI: 10.1021/acs.chemmater.1c01071

Latest Updates

See All