2008 APS March Meeting
Volume 53, Number 2
Monday–Friday, March 10–14, 2008;
New Orleans, Louisiana
Session H4: Selected Applications Using Materials Science
8:00 AM–11:00 AM,
Tuesday, March 11, 2008
Morial Convention Center
Room: 206
Sponsoring
Unit:
DMP
Chair: Yvan Bruynseraede, Katholieke Universiteit Leuven
Abstract ID: BAPS.2008.MAR.H4.5
Abstract: H4.00005 : Materials Informatics: Using machine learning techniques with large amounts of ab-initio computed or experimental data
10:24 AM–11:00 AM
Preview Abstract
Abstract
Author:
Gerbrand Ceder
(Massachusetts Institute of Technology)
Machine learning techniques can be applied to large amounts
experimental or
computed materials data in order to identify the underlying
factors that
determine a target property. While the use of experimental data is
complicated by the fact that it is mostly non-standardized in
property or
structure databases, experimental data still tends to be richer in
information than computed data. One problem that can be addressed
with
machine learning techniques is the prediction of structure. By using
structure prototype as a mathematical descriptor, and
constructing its
correlation in chemical spaces through machine learning
techniques, it is
possible to create a highly effective structure prediction method.
Previously, we demonstrated that by simply applying maximum
entropy ideas to
a large experimental structure database of binary metals, it was
possible to
suggest a short list of candidate structures for new compounds which
contains the proper ground state with very high probability
[Ref]. This list
of probable structures can then be computed with ab initio energy
methods.
We have now extended this method to multi-component and non-metal
systems by
prototyping the $\approx $ 100,000 structure records in the
International
Crytallographic Structure Database, and a similar accuracy of
prediction is
achieved in these high component spaces. We believe that such a
machine
learning approach solves the crystal structure prediction for
many practical
purposes. Machine learning techniques can also be used to point
at likely
errors in experimental structure databases and I will give some
examples of
this.
In the long-term computed data is more likely to form the input
for machine
learning techniques as it is well defined and obtained under
controlled
conditions. Using high-throughput ab-initio computing techniques
we have
determined the structure and energy for several thousand
compounds and have
begun to data mine this information for property models relevant
to energy
generation and storage.
To cite this abstract, use the following reference: http://meetings.aps.org/link/BAPS.2008.MAR.H4.5