Forum Schedule Fall 2020

Kinetics plays an important role in both performance and durability of functional materials. For rechargeable batteries specifically, charge/discharge requires fast kinetics to meet application needs, and meanwhile side reactions have to be slow for safety and durability. To achieve the faster-charging and longer-lasting goal, it is essential to understand the related kinetic processes on the atomistic level.

In this talk, I will discuss rate-limiting steps in cathode materials by combining density functional theory and transition state theory and demonstrate how these atomistic pictures help accelerate Li transport and suppress degradation. The first part will be on Li transport in LiFePO₄. Two rate-limiting steps are identified for extracting Li from LiFePO₄: one is the Li-poor phase nucleation in the bulk and the other is Li passing through the particle surface. Surface sulfur deposition is proposed and confirmed as a new approach to improve the rate capability. The second part will be on degradation in layered oxides. One degradation channel is the oxygen loss at the highly charged state. This process starts at particle surfaces, and thus can be kinetically hindered by stabilizing surface oxygen. Our computational results show that forming SO₄ groups on oxide surfaces alleviates oxygen loss and electrolyte oxidation. Borrowing the same sulfur deposition procedure from the first part, we confirm experimentally that surface SO₄ improves the cyclability of a Li-rich layered oxide. The above understandings on Li transport and degradation kinetics could help rational design of better rechargeable batteries.

The recent years have seen a surge in the development of machine learning algorithms in different areas of scientific research. In the field of simulation of condensed matter, the development of machine learning force fields to carry out molecular dynamics simulations, is a topic that has attracted a lot of interest ever since the pioneering work of Behler and Parrinello in 2007.

Machine learning force fields are trained using reference data coming from expensive ab initio simulations, and try to approximate these accurate methods in a computationally more efficient way. Many methods have been developed and showcase good force and global energy prediction accuracies with respect to the data used to train them, usually obtained from ab initio calculations, e.g. density functional theory (DFT) calculations.

After a brief introduction regarding the differences between artificial neural networks and Bayesian methods such as Gaussian process regression, we present the case for the use of Gaussian process regression to construct machine learning force fields that retain the interpretability of classical parametrized force fields, but are inherently non-parametric and can, therefore, be automatically trained from reference data. [1]

We design algorithms to construct force fields that have an explicit dependence on atomic pair distances (2- body), angles (3-body) and atomic embedding (many-body). Then, by exploiting this explicit dependencies, we are able to ``map” the Gaussian process force fields into nonparametric classical force fields, thus increasing very substantially the computational speed in prediction [2, 3].

These ``mapped” force fields have been recently employed for MD simulations of Ni19 nanoparticles [4], and for a variety of bulk materials [5], and ongoing research aims to increase their accuracy in challenging systems such as Au nanoparticles and metallic surfaces for catalytic reactions.

[1] Glielmo, A., Sollich, P., De Vita, A. (2017). Accurate interatomic force fields via machine learning with covariant kernels, Physical Review B, 95(21), 214302

[2] Glielmo, A., Zeni, C., & De Vita, A. (2018). Efficient nonparametric n-body force fields from machine learning. Physical Review B, 97(18), 184307.

[3] Glielmo, A., Zeni C., Fekete, Á., De Vita A. Building nonparametric n-body force fields using Gaussian process regression. In: Machine Learning for Quantum Simulations of Molecules and Materials, 1st ed. (accepted manuscript)

[4] Zeni, C., Rossi, K., Glielmo, A., Fekete, Á., Gaston, N., Baletto, F., & De Vita, A. (2018). Building machine learning force fields for nanoclusters. The Journal of Chemical Physics, 148(24), 241739.

[5] Vandermause, J., Torrisi, S. B., Batzner, S., Kolpak, A. M., & Kozinsky, B. (2019). On-the-fly Bayesian active learning of interpretable force-fields for atomistic rare events. arXiv preprint arXiv:1904.02042.

As the saying goes, seeing is believing. Visualizing molecules and their properties is essential for enhancing our understanding of chemistry and molecular interactions. I will describe two very different applications of molecular visualization. In the first, I will describe the 3Dmol.js Javascript library. 3Dmol.js allows users to embed fully interactive, 3D accelerated molecular structure views into web pages and Jupyter notebooks. It also supports a hosted active learning environment that allows teachers to query students about molecular structures in real-time. I will provide examples of using 3Dmol.js for education and research and conduct an interactive demonstration of the learning environment.

In the second, very different, application of visualization to understanding I will describe our grid-based neural network potentials for protein-ligand scoring and other molecular properties. I will show how we can attempt to crack open the black box of deep learning to understand how deep neural networks make predictions through various visualization schemes.