The primary ML method in this project has been high-dimensional neural network potentials (HDNNPs). At the start of the project, there already existed several open-source packages which implement this particular method. One package in particular, TorchANI, stood out as the package distributors included with pre-made, generally applicable, force fields. Within this project, the fellow integrated the code into the Amsterdam Modeling Suite (AMS), allowing users to get DFT or CC-quality results for organic molecules almost instantly. Furthermore,
the integration of TorchANI was accomplished via a general and flexible interface that also supports other ML Python packages for predicting potential energy surfaces (such as SchNetPack, sGDML, and PiNN), allowing for many different types of ML methods to be used with AMS.
In a collaboration with researchers from Uppsala University, the fellow developed PiNN, an open-source Python package for constructing and evaluating atomic neural networks for molecules and materials. This package implements not only the high-dimensional neural network potentials, but also message-passing graph-convolutional neural networks, in particular the PiNet architecture which was developed in this project. One benefit of this approach, in comparison to HDNNPs, is that the features of the local atomic environments do not need to be constructed explicitly beforehand, but these features are instead learned by the ML method. The PiNet architecture can be parameterized to accurately predict the potential energy surface and its gradients (forces, and stress tensors for periodic systems). Moreover, it can be applied to directly predict properties, for example the formation energies of materials.
As a result of this project, it is now possible to run simulations inside the Amsterdam Modeling Suite combining a variety of low-level methods like ReaxFF, DFTB, or DFT with Machine Learning to improve the original predictions. The interface is very general and also allows to perform quantum-mechanics/molecular-mechanics (QM/MM) hybrid calculations. It is also possible to parametrize the corresponding ML methods.
Although the focus of this project was on Machine Learning, the fellow has also deepened his knowledge about other parametrized methods, in particular ReaxFF. ReaxFF is a reactive force field with many parameters that need to be fitted. The fellow co-developed ParAMS, a Python package for fitting parameters for any of the many methods implemented in AMS, including ReaxFF. ParAMS handles training set and validation set evaluations, features a variety of fitting algorithms, and has a special focus on transparency and reproducibility.
These new developments have been released as part of SCM’s Amsterdam Modeling Suite 2020, and are thus available to the materials modelling community. Likewise, the PiNN Python library is available on GitHub.
Finally, the fellow was involved in two more applied collaborations with experimental researchers. In the first, ML simulations were used to elucidate temperature effects on the ionic conductivity in concentrated alkaline electrolytes, in particular how deviations from the Nernst-Einstein theory of conductivity are amplified by proton transfer reactions. In the second collaboration, the fellow’s ML simulations and DFT calculations were used to elucidate the formation mechanism of a particular zeolite in very alkaline silica solutions. By combining these simulations with state-of-the-art experimental techniques, the collaborative effort could for the first time illustrate how an inorganic ion-pair could be a structure-directing agent during zeolite synthesis.
