Machine Learned Predictions of Complex Quantities from Differentiable Networks

The use of machine learning methods in condensed matter simulation presents some advantages in comparison to ab initio methods. Notably, using a trained model to calculate properties of a system can often be orders of magnitude faster than doing a DFT calculation, with a similar level of accuracy. However, a significant amount of data must be generated beforehand, which can cancel this advantage, especially when studying more complex quantities, such as vibrational properties and Raman spectra.

By taking advantage of the highly differentiable architecture of neural networks, we developed a package¹ allowing direct predictions of the derivatives of the quantities present in the training data. In the cases of derivatives with respect to atomic positions, this requires calculations of out of equilibrium structures. We are working on a method to optimize the data generation of these structures and the training of models in a single fully machine learned workflow, aiming to reduce the number of data points needed and the biases they carry.

1. https://github.com/OMalenfantThuot/ML_Calc_Driver