About CPMD
CPMD meetings have a tradition stretching back two decades in bringing together a diverse mix of computational scientists working on different aspects of molecular dynamics simulations, combining aspects of electronic structure theory and statistical sampling methods, as well as state-of-the-art applications ranging from materials science to biophysics. In this, they very much target a core area of interest for Psi-k, bridging statistical mechanics and quantum mechanics.
This edition of the meeting - which marks the 20th anniversary of the first CPMD meeting at Ringberg Castle (Munich) - will focus on identifying the most pressing challenges in the field, and on discussing the most promising directions to face such challenges. It is quite clear that one of the most promising, potentially disrupting developments over the past few years involves the use of machine-learning methods in the field of atomic-scale modelling. This topic will be a particular focal point for our selection of speakers and the organization of the program. However, we will not cover exclusively machine-learning techniques - that have been extensively discussed by a multitude of workshops and conferences over the past few years. Rather, we will build on the traditional breadth of interests of this community, and consider machine learning in the broader context of the challenges faced by molecular dynamics simulations, considering three distinct perspectives:
- Pushing the boundaries for electronic structure methods in ab initio molecular dynamics. We will cover simulations that use beyond-DFT methods, and high-performance computing optimization to enable more accurate and/or lower-cost electronic structure calculations to be used in AIMD simulations. With the advent of machine-learning potentials, it becomes even more crucial to provide very accurate reference values, that can be used not only for direct sampling, but also to train surrogate quantum models.
- Data-based science and machine learning. While we want to consider machine-learning in the context of molecular dynamics, we obviously cannot overlook the process of building a machine-learning model. We will ask speakers discussing predominantly ML algorithms to present their work including a broad perspective, highlighting how it challenges several of the fundamental assumptions underlying molecular dynamics simulations of materials, molecules and biological systems. We will discuss how these developments change the landscape of molecular modelling -- including the development of machine-learning potentials (that links directly to the first topic), the design of optimal descriptors to predict material properties, and the automatic determination of order parameters for sampling (which is very relevant for the third perspective we want to cover).
- Pushing the boundaries for sampling, and long time-scale modelling. We will discuss different approaches - not only MD-based - to compute more accurate free-energies for both stable phases, reactive events, and conformational changes in proteins, including effective thermostatting schemes, replica exchange methods, nested sampling, and biased sampling algorithms. We will suggest the speakers to emphasize unbiased, data-driven approaches to build order parameters and optimize sampling, as well as to discuss how sampling can be used to generate a more comprehensive training set for machine learning.
Speakers will be asked to present at least one slide highlighting their vision on the directions of the field, and we will close the meeting with a panel discussion summarizing the emerging trends. This discussion will be summarized in a short whitepaper that will be included in the PSI-k report.
Organizers
Michele Ceriotti (EPFL)
Nicola Marzari (EPFL)
Matteo dal Peraro (EPFL)
Alfredo Pasquarello (EPFL)
