[Back]

B. Hartl, G. Kahl:
"Nested Effective Potentials";
accepted as talk for: Machine Learning and Reverse Engineering for Soft Materials, Leiden, Netherlands; 12-10-2018 - 12-14-2018; in: "Machine Learning and Reverse Engineering for Soft Materials", (2018).

The emergence of machine learning seems to offer great support in a vast number of scientific and technological fields. In computational physics, for instance, machine learning approaches increasingly establish themselves as efficient tools in force-field-modeling or classification of atomistic or molecular systems.
By using neural-network-potentials [1, 2] it has been shown that simulations of water can be carried out on molecular-dynamics time-scales with density-functional-theory (DFT) precision [3].
The main task here is to provide a (large) training-data-set -- capturing the essential physics -- during the learning stage of the neural-network.
The latter then effectively performs force-field-fitting to the training-data such that the network eventually models the physics of the system.

Neural-network potentials are not restricted to particle number, however, their applicability to heterogeneous systems is limited [2].
Therefore it currently seems difficult to apply this strategy to heterogeneous systems involving large numbers of complex molecules (being composed of functional groups with atomistic components).
For such systems an atomistic treatment is rather difficult and simulations or the search for their ground-states is time consuming and often not feasible.
Approximations and coarse-graining approaches sometimes miss essential parts of the physics like asymmetry, deformability or the possibility of chemical reactions.

The question I want to address in this proposal is, if it is possible to generate a hierarchical coarse-graining of complex molecular systems based on machine-learning potentials,
starting from atomistic (e.g. DFT-based) data, while maintaining -- as good as possible -- the atomistic accuracy?

To be more specific:

- How does a machine-learning system need to be setup to identify essential parts of a molecule and provide effective potentials for these building-blocks?
- Are these machine-learning potentials good candidates for further coarse-graining and how well do these hierarchical potentials fit the most atomistic one?
- Is it possible to identify a strategy how these building-blocks need to be attached to each other in order to form reasonable molecules?
- Can these building-blocks be designed in order to allow chemical reactions, that is to change their type during simulation?

The model/algorithm would identify groups of atoms as building-blocks, groups of building-blocks as molecules and so on, and would be able to provide force-fields at any level of resolution.
The hierarchic nature of this approach would greatly reduce computational costs (again on every level of resolution) without losing (too much) accuracy.
The building-block type of treatment would intrinsically allow for scalability and could be hosted as a database which would be valuable for the entire community.
Eventually technological applications could greatly benefit from such models.

[1] S. Lorenz, A. Groß, and M. Scheffler, Chem. Phys. Lett. 395, 210 (2004).
[2] J. Behler 2014 J. Phys.: Condens. Matter 26 183001 (2014).
[3] T. Morawietz, A. Singraber, C. Dellago, and J. Behler, PNAS 113 (30) 8368-8373 (2016).

Project Head Benedikt Hartl:
Soft and Anisometric Self-Assembly

Project Head Gerhard Kahl:
DFS