Menu
A+ A A-

Webinar: Designing Novel Nanoporous Materials for Applications in Energy and Environment. From Multi-Scale Modeling to Materials Informatics

Event Details:

  • Date:          Tuesday, 6 April 2021
  • Time:         Starts: 16:00
  • Venue:        Live streaming of the discussion will be available on Zoom (Password: VsSCz1).
  • Speaker:  Prof. George E. Froudakis, Department of Chemistry, University of Crete, Greece



CyI Logo RTI ver b     CaSToRC HPC

 
CaSToRC, the HPC National Competence Centre,
 invites you to the EuroCC and SimEA Online Seminar Series

 
The webinar will be in English and the live stream is open to the public.
Live streaming of the discussion will be available on Zoom (Password: VsSCz1).
Images and/or recordings of our open public events may be used by The Cyprus Institute for dissemination purposes including print and digital media such as websites, press-releases, social media, and live streaming.

 


                                                                                                             

Abstract

Machine learning techniques (ML) are powerful tools already used in science and industry since their computational cost is by several orders of magnitude lower than that of the “conventional” approaches. However, their ability to provide accurate predictions strongly depends on the correct identification of those parameters (descriptors) that will allow the algorithm to effectively learn from past data. Other critical factors that affect the quality of the predictions are the size and the quality of the dataset used for the training of the algorithm as well as the correct estimation of the training size.

Aiming at both, the transferability of our model and the reduction of the training data set, we introduce 2 different classes of descriptors, based on fundamental chemical and physical properties: Atom Types and Atom Probes. The main difference from previous models is that our descriptors are based on the chemical character of the atoms which consist of the skeleton of the materials and not their general structural characteristics. With this bottom up approach we go one step down in the size of the descriptors employing chemical intuition.

On parallel, an automatic procedure of identifying the appropriate size of the training set for a given accuracy was developed. A novel training algorithm based on “Self-Consistency” (SC) replaced the standard procedure of linearly increasing of the training set. Our SC-ML methodology was tested in 5.000 experimentally made MOFs for investigating the storage of various gases (H2, CH4, CO2, H2S, H2O). For all gases examined, the SC-ML methodology leads to significantly more accurate predictions, while the number of MOFs needed for the training of the ML algorithm in order to achieve a specified accuracy can be reduced by an order of magnitude. In addition, the universality and transferability of our ML model was proved by predicting the gas adsorption properties of a different family of materials (COFs) after training of the ML algorithm in MOFs.

Despite the progress in the field and the improved models that have been recently developed, ML algorithms fail to classify new materials with improved properties compared to the known ones. To the best of our knowledge, the previous point has never been addressed since extrapolation is an inherent drawback of ML. The reason behind this drawback is mainly attributed to the fact that for reliable predictions of top-performing materials (materials with very high gas-adsorption capacities), ML algorithms need to be trained using materials of the same or higher performance.

In lack of such information, we propose a new methodology for the construction of artificial data (artificial MOFs) with the desired properties that will be used for ML training. This will enable ML algorithms for achieving improved predictions, in particular for high-performing materials. We demonstrate that, after using the artificial data, the capability of the ML algorithms to classify new top-performing MOFs as such, improves remarkably. We are also confident that the present methodology represents an important contribution toward the development of predictive models aiming to the discovery of new materials with outstanding properties. In addition, the main idea of this approach, can be used in many other applications of ML methodologies for overcoming the inherent problem of extrapolation. 

About the Speaker

George FroudakisProf. George E. Froudakis

Department of the University of Crete. He is the author of more than 150 publications, is on the board of directors of the Greek Hydrogen Platform and a Founding member of the Greek National Science & Innovation Foundation. He has coordinated 10 European and Greek research projects and participated in 16 more. He has supervised 6 Postdocs, 10 PhD and 12 MSc.

His research activities are focused on designing, modelling and investigating properties of nanostructures and porous materials for Energy, Environment and Health applications. Multi-scale computational techniques are developed in-house and used for simulating large systems. Recently, a new computational methodology for large-scale screening of materials with the use of Machine Learning algorithms (ML) was developed.

 

 

Download the Spring 2021 Online EuroCC & SimEA Seminar Series Programme here.


EuroCC Simea Seminar Series 2
 



 

Contact 
This email address is being protected from spambots. You need JavaScript enabled to view it.

 

View all CyI events.

 

Additional Info

  • Date: Tuesday, 6 April 2021
  • Time: Starts: 16:00
  • Speaker: Prof. George E. Froudakis, Department of Chemistry, University of Crete, Greece