Investigation of chemical reactivity by machine learning techniques

The prediction of chemical processes and materials properties requires accurate knowledge of potential energy surfaces. These map atomic positions and nuclear charges to electronic energies. Such potential energy surfaces can be obtained point-wise by electronic structure calculations. Based on these, machine learned models can be trained if appropriate descriptors of atomic structures are available. They have to fulfill physical constraints, like invariance with respect to rotation, translation, and reflection of the whole molecule of invariance with respect to exchange of atoms of the same type. Such descriptors are to be developed in this project and will be applied in ML approaches like feed-forward neural networks and kernel ridge regression. Furthermore, only a single neural network with a square augmented input layer will be used for all atomic species, which is different from the most available open-source packages. In the case of the kernel ridge regression model an atomic selectivity of the kernel will be introduced. Along with standard kernels, like square exponential one, different graph kernels will be used.