CSE8803 / ME 8883 Materials Informatics
Fall 2016 Georgia Tech
Anh Tran and Zhiyu Wang
- High-dimensional PES metamodel via DFT - Blog 5: Battling against the curse of dimensionality using permutation in material science (Group:WangTran)
- High-dimensional PES metamodel via DFT - Blog 3: Reduce the curse of dimensionality on meta-modeling (Group:WangTran)
- High-dimensional PES metamodel via DFT - Blog 2: High-dimensional kriging with modifications (Group:WangTran) - Report 21Sep16
- High-dimensional PES metamodel via DFT - Blog 1: Data acquisition (Group:WangTran) - Report 13Sep2016
- High-dimensional PES metamodel via DFT - Blog 7: Simulating and searching for the MEP on PES (Group:WangTran)
- High-dimensional PES metamodel via DFT - Blog 4: Searching for local minima and saddle points algorithm with kriging (Group: WangTran)
- High-dimensional PES metamodel via DFT - Blog 8: Searching and refining based on Kriging metamodel (Group:WangTran)
- High-dimensional PES metamodel via DFT - Blog 9: The limitations of metamodel (Group:WangTran) - Report 19Nov2016
- High-dimensional PES metamodel via DFT - Blog 10: Final Report - Constructing high-dimensional potential energy surface via metamodel
- High-dimensional PES metamodel via DFT - Front Page
Today we are concerned with the searching algorithm for the Fe8H system. The FeTiH system is similar, but any of them take gigantic computational time. It has been 10 days to run on the PACE cluster and waiting for the results to come back. Meanwhile, we have some data to do post-process analytics. As far as I know, the surface is (very) rough and PCA does not represent them well, as regression cannot be applied for potential surface with high accuracy.
The searching algorithm is pre-designed by one of our former Ph.D. students, Lijuan He, and we are taking it to another stage where it can be applied to quantum chemistry, where high-dimensionality poses a challenge. The algorithm is constructed in such a way that it can locate the local minima and solve for the minimum energy path (MEP), given the meta-model of the potential energy surface (PES). This searching algorithm performs well under numerical testing functions in optimization, such as Rastrigin and Schwefel, but remains incomplete in real material system. As proposed before, we are approaching the problem using the kriging technique. We have written a code to sort the Fe atom in Fe8H. Potentially in the future, we can use clustered kriging – a modified technique for kriging to approach exact interpolation with big data.
Experimental validation include 2 sites of possible H sites, one octahedral and one tetrahedral. We should expect the numerical simulation agree with the experimental results, to some degree, if not perfectly. This is the result that we keep in mind in the long run of this study.
Figure 1 shows the interstitial sites in a body-centered lattice with metal atoms as big black dots. Figure 1(a) shows the octahedral sites with small blue dots, where Figure 1(b) shows the tetrahedral site with small blue dots. Those meta-stable states representing by local minima on PES are the experimental results that we aim to validate our searching algorithm.