Machine learning predicts defects in materials

Machine learning approaches in materials have focused on the prediction of materials with new properties. However, realistic materials exhibit defects. The natural question is how defects affect the properties of the predicted material? To tackle this question for magnetic materials, Prof. Iacocca’s group has developed a statistical pseudospectral Landau-Lifshitz equation of motion that takes defects into account by its spectral response. Because defects can be randomly located and, in principle, have random dimensions, their distribution is well represented by a random telegraph noise.

This model can be readily integrated into deep learning algorithms. In a recent publication in npj Computational Materials, the authors demonstrated deep learning methods to predict the dispersion relation and the domain-wall width for a material with strong perpendicular magnetic anisotropy. For the dispersion relation, a convolution neural network and a surrogate model built by using theory of functional connections was able to reproduce the dispersion relation bounded by physical constraints. For the domain-wall width, a convolutional neural network augmented with the defect parameters was able to predict domain-wall widths with a standard deviation of two atomic lattices. These results are expected to pave the way for more accurate method that leverage statistical approaches to predict physical limits of new functional materials.