Frontiers in Molecular Dynamics: Machine Learning, Deep Learning and Coarse Graining
- Amir Natan (Tel Aviv University, Israel)
- Yair Shokef (Tel Aviv University, Israel)
- Rickard Armiento (Linköping University, Sweden)
The workshop happens on October 10th-12th, 2018.
Prior to the workshop, there would be a 2 days school on October 8th-9th, 2018.
The calculation of the atomistic dynamics of large systems is an important challenge in many fields. The use of classical force fields (FF) [1-2], where specific physical models are built to describe the interaction between atoms and molecules, is a very successful approach that allows a realistic description of forces in large systems at a very low computational cost. While such FF models have been very useful to describe large systems they have several limitations – first - the model for the forces might not be accurate at all configurations, second – chemical reactions in the system or radical changes in the atomic environment are almost impossible to describe with a simple classical FF. In recent years, more sophisticated FF schemes, such as variable charge FF , ReaxFF [4-5], COMB [6-7], and others, have appeared. Such FF include a very large set of parameters that can successfully describe a much richer set of cases, this approach was successful in describing complicated system such as surface oxidation, chemical reactions and many others. The limitation of the more sophisticated force fields is that they are still tailored to a limited set of configurations (although much wider) and in addition, a heavy parameterization is now needed. A different approach is building “on the fly” schemes for forces estimation, such schemes use quantum calculations as a training set to estimate the forces. Machine Learning (ML) [8-10] and Deep Learning (DL) [11-13] were shown in recent years to successfully predict energies, forces and electronic properties with a reasonable size datasets that are updated on the fly. A good prediction of forces can lead to an ab-initio accuracy molecular dynamics (MD) at a computational cost that although higher than classical FF is still much faster than fully ab-initio MD. Several groups have achieved such implementations and although this is not in common use of general purpose MD simulations, there is a great promise for future applications and a growing community. Another important aspect of modern MD is coarse graining, enabling to reach even larger systems and to approach mesoscopic and continuum scales. While this seems like a completely different area, there might be future interaction as the modeling of long range forces and coarse graining are somewhat related.
The goal of this 2.5days workshop is to have a meeting of some of the leading researchers in the field to discuss the impact of both machine learning and deep learning on molecular dynamics techniques. We invite also researchers that develop coarse graining techniques and large scale soft matter simulations. The workshop will include lectures of the researchers, discussion panels and a poster session. Prior to the workshop we intend to have a 2.5days school with more basic lectures in:
• Machine Learning and Deep Learning in the context of materials research and molecular dynamics.
• Advanced coarse graining methods.
• QM/MM techniques.
In the workshop we would like to compare state of the art machine learning and deep learning approaches. We would also like to check how close are those methods to be able to simulate large and complicated systems.
We have invited some of the leading researchers in this field and plan to invite additional researchers that do ML and DL in materials science.
 Understanding Molecular Simulation: From Algorithms to Applications; Daan Frenkel, Berend Smit, Academic Press 2001
 Computer Simulation of Liquids, M. P. Allen, D. J. Tildesley, Clarendon Press, 1989
 Streitz, F. H., and J. W. Mintmire. "Electrostatic potentials for metal-oxide surfaces and interfaces." Physical Review B 50, no. 16 (1994): 11996.
 Van Duin, Adri CT, Siddharth Dasgupta, Francois Lorant, and William A. Goddard. "ReaxFF: a reactive force field for hydrocarbons." The Journal of Physical Chemistry A 105, no. 41 (2001): 9396-9409
 Aryanpour, Masoud, Adri CT van Duin, and James D. Kubicki. "Development of a Reactive Force Field for Iron− Oxyhydroxide Systems." The Journal of Physical Chemistry A 114, no. 21 (2010): 6298-6307.
 Liang, Tao, Bryce Devine, Simon R. Phillpot, and Susan B. Sinnott. "Variable charge reactive potential for hydrocarbons to simulate organic-copper interactions." The Journal of Physical Chemistry A 116, no. 30 (2012): 7976-7991
 Cheng, Yu-Ting, Tzu-Ray Shan, Tao Liang, Rakesh K. Behera, Simon R. Phillpot, and Susan B. Sinnott. "A charge optimized many-body (comb) potential for titanium and titania." Journal of Physics: Condensed Matter 26, no. 31 (2014): 315007
 Rupp, Matthias, Alexandre Tkatchenko, Klaus-Robert Müller, and O. Anatole Von Lilienfeld. "Fast and accurate modeling of molecular atomization energies with machine learning." Physical review letters 108, no. 5 (2012): 058301.
 Botu, Venkatesh, and Rampi Ramprasad. "Adaptive machine learning framework to accelerate ab initio molecular dynamics." International Journal of Quantum Chemistry 115, no. 16 (2015): 1074-1083.
 Li, Zhenwei, James R. Kermode, and Alessandro De Vita. "Molecular dynamics with on-the-fly machine learning of quantum-mechanical forces." Physical Review Letters 114, no. 9 (2015): 096405.
 Behler, Jörg, and Michele Parrinello. "Generalized neural-network representation of high-dimensional potential-energy surfaces." Physical review letters 98, no. 14 (2007): 146401.
 Behler, Jörg. "High‐Dimensional Neural Network Potentials for Complex Systems." Angewandte Chemie International Edition (2017).
 Schütt, Kristof T., Pieter-Jan Kindermans, Huziel E. Sauceda, Stefan Chmiela, Alexandre Tkatchenko, and Klaus-Robert Müller. "MolecuLeNet: A continuous-filter convolutional neural network for modeling quantum interactions." arXiv preprint arXiv:1706.08566 (2017).