This is the second session of the second season of the Mixed-Gen on-line series aimed mainly at PhD students and researchers in their first post-doc. Our goal is to continue providing a new venue for these young scientists to share their work, get expert feedback and have an opportunity to strengthen scientific relations within the CECAM community.

The general area for this session is Machine learning in simulations

To participate

If you are a PhD student or a post-doc:

Please use the Participate Tab on this page to start the application. You will have to login using your CECAM account to access the application form. If you don't have a CECAM account yet, use the register option on the top right corner of the login page...and welcome to CECAM!

If you are a more senior scientist:

Please contact the organisers and we shall process your registration.

Submission of posters

(Please note that - at least for the time being - we shall accept posters only from PhD students or researchers in their first post-doc)

After your application is accepted, you will be able to submit a poster. In the CECAM page for this event, go to “My participation” tab and click on “Add a poster”, providing title and abstract following the recommended format. On the same form you can upload your poster file in png or jpg as soon as it is ready. These formats are strict to enable showing of the poster in the Gather session. If the poster file is not ready at the moment of submitting your abstract, you can upload it later by editing your submission (Go to “My participation” tab and click three vertical dots on “Actions” column on table “My posters”). Please upload your poster as soon as possible to enable a decision from the selection committee - see below.

Please note that posters will be visible in the Gather room associated with this session until the end of the series (June 2022) unless otherwise requested.

DEADLINE FOR SUBMISSION: TEN DAYS BEFORE THE EVENT

Selection of posters

Posters will be selected by the event organisers with the support of our main speaker and experts who will take place in the poster session. 

Selection of the two talks by PhD or first year postdocs

These contributions, to be broadcasted in the Zoom webinar in the first part of the event, will be selected, after a preliminary screening by the organisers, the main speaker and guest experts, via a lottery from the posters selected for the Gather session. Please tick “No” to the question “Upgrade to talk?” in your application if you DO NOT WANT your poster to be considered for this lottery. 

THE DECISION ON THE POSTER AND THE OUTCOME OF THE LOTTERY SELECTION WILL BE COMMUNICATED ONE WEEK BEFORE THE EVENT

POSTER SUBMISSIONS BEYOND THIS DEADLINE WILL BE ACCEPTED BUT NOT CONSIDERED FOR UPGRADE TO TALK. SUBMISSION WILL BE DEFINITELY CLOSED FOUR DAYS BEFORE THE EVENT.

SESSION 2. Title and abstract of talks

First principles force fields: status and challenges

Gabor Csanyi, Cambridge University

Over the past decade a revolution has taken place in how we do large scale molecular dynamics. While previously first principles accuracy was solely the purview of explicit electronic structure methods such as density functional theory, the new approaches have allowed the extension of highly accurate, first principles simulations to the atomic scale, where electrons are not treated explicitly any more, and therefore hundreds of thousands of atoms can be simulated. These quantum mechanically accurate force fields and interatomic potentials are fitted to electronic structure data and at first used techniques inspired by those used in machine learning and artificial intelligence research: neural networks, kernel regression, etc. It is a quickly moving field, and - having learned key lessons about representation, symmetry and regularisation - there appears to be some semblance of convergence between the diverse methods, which now also include polynomial expansions carried to high dimension. 

Using machine learning to improve RNA force fields

Thorben Fröhlking, Giovanni Bussi, SISSA

The capability of current force fields to reproduce structural dynamics in agreement with experiment is limited. In the past years, several methods have been developed to take advantage of experimental data in order to enforce agreement with experiments. Since Maximum-entropy methods are limited with respect to the choice of the corrections functions, the approach introduced in Cesari et al., J. Chem. Theory Comput. 2019, 15, 6 provided a more flexible method. This approach allows arbitrary force-field terms to be corrected, and arbitrary ways to quantify agreement with experiment, together with a robust regularization protocol to avoid overfitting. We here extend this framework introducing and comparing a number of different regularization strategies (L1, L2, Kish Size, Relative Kish Size, Relative Entropy). The protocol is applied to difficult RNA systems, namely GACC, ccGAGAgg and ccUUCGgg, where specific intramolecular hydrogen bonds in the AMBER RNA force field proposed by Kührová, et al., J. Chem. Theory Comput. 2019, 15, 5 are corrected with automatically determined parameters.  The identified parameters are in agreement with recently suggested corrections Mráziková et al., J. Chem. Theory Comput. 2020, 16, 12 as well as the experimental NMR data. Simulations involving systems present in the training set, but also unknown systems, using identified parameters display improvements regarding native population of Tetraloops as well as good agreement with NMR-experiments for Tetramers.

Furthermore, we test the possibility to introduce non-linear corrections in the form of two-layers artificial neural networks (ANN) applied on pairs of consecutive dihedral angles. By testing on an extended dataset, we find that the linear corrections on dihedral angles introduced in Cesari et al., J. Chem. Theory Comput. 2019, 15, 6 as well as introduced ANN potentials significantly reduce existing discrepancies between experiments and simulations regarding NMR observables.

References

[1] A. Cesari, S. Bottaro, K. Lindorff-Larsen, P. Banáš, J. Šponer, G. Bussi, J. Chem. Theory Comput., 15, 3425-3431 (2019)
[2] T. Fröhlking, M. Bernetti, N. Calonaci, G. Bussi, J. Chem. Phys., 152, 230902 (2020)
[3] P. Kührová, V. Mlýnský, M. Zgarbová, M. Krepl, G. Bussi, R. Best, M. Otyepka, J. Šponer, P. Banáš, J. Chem. Theory Comput., 15, 3288-3305 (2019)

Machine Learning of Metamaterials

Silvia Bonfanti, Università degli Studi di Milano

Mechanical metamaterials are a new class of materials with exceptional properties derived from their designed internal geometry composed of multiple sub-elements or cells, rather than their composition. Through prudent choice of cells arrangement, mechanical metamaterials find a straightforward application as actuators that achieve predetermined input-output operations and can be 3D printed in a single block, with the advantage of removing the assembly phase since there is no longer need for all structural components. Despite the rapid progress in the field, there is still a need for efficient strategies to optimize metamaterial design for a variety of functions.

Here, we first present a computational method for the automatic design of mechanical metamaterial actuators that combines a reinforced Monte Carlo algorithm with discrete element simulations. The method allow us to get machine-generated structures that can reach high efficiency in terms of functionality, exceeding human-designed structures.

Next, we show that it is possible to design novel efficient actuators by training a deep neural network which is then able to predict the efficiency from the image of a structure, and to identify its functional regions which are those mechanically relevant.

The use of machine learning to assist the automatic design of mechanical metamaterial actuators opens intriguing possibilities in terms of algorithmic speed, as it could potentially allow to design larger structures that can not be efficiently simulated by conventional methods, increasing the complexity for countless engineering applications.