Deep learning crystal field parameters from thermodynamic observables

Visual of workflow from CNN input wavelet to crystal field parameter output.Scientific Achievement

Discovery of deep neural net algorithm that extracts crystal field parameters in an efficient and unbiased way from thermodynamic results.

Significance and Impact

Demonstration that deep learning  accelerates material characterization.

Research Details
  • Input of convolutional neural network (CNN) is thermodynamic data like specific heat and magnetization
  • Wavelet transform is used to transform raw data into an image that is recognized by the neural net. 
  • Can be applied to magnets from the full rare-earth series: benchmark results for CeAgSb2, PrAgSb2 and PrMg2Cu9.

N. F. Berthusen, Y. Sizyuk, M. S. Scheurer, P. P Orth, "Learning crystal field parametersusing convolutional neural networks." SciPost Phys. 11, 011 (2021).