Augmenting DFT modelling of vibrational spectra

Augmenting DFT modelling of vibrational spectra through Machine Learning and Deep Neural networks using small data sets.

Within this project we aim to accelerate the quantum mechanical prediction of vibrational spectra allowing for their use in large scale complex systems.

Vibrational spectra provide unique fingerprints of molecules and diamond color centers, making them a powerful tool for materials characterization. However, obtaining these spectra from first-principles calculations is computationally expensive, as it requires the evaluation of the dynamical matrix using finite-difference density functional theory (DFT).

In this project, we explore the use of machine learning to predict dynamical matrices and thereby reduce the computational cost of vibrational calculations. Graph neural networks are trained on relatively small datasets of compact systems, with the goal of extrapolating to larger and more realistic structures. A key challenge is that the dynamical matrix is rotationally equivariant, meaning that suitable machine-learning models must explicitly respect this symmetry. By incorporating equivariance into the model design, we aim to enable accurate and efficient predictions of vibrational properties in complex systems.

Thijs van Wijk, MSc. Physics’23