Download MACHINE LEARNING OF SURFACE ADSORBATE STRUCTURE M

MACHINE LEARNING OF SURFACE ADSORBATE STRUCTURE
M. Todorović1 , M. U. Gutmann2 , J. Corander2 and P. Rinke1
1
Aalto University, P.O. Box 11100, FIN-00076, Espoo, Finland
University of Helsinki, P.O. Box 68, FIN-00014, Helsinki, Finland
email: [email protected]
2
The adsorption and self-organisation of molecules at inorganic surfaces is central to many
industrial processes from catalysis and coatings, to organic electronics and solar cells.
Since structure determines function, any computational study of pertinent processes first
requires knowledge of the interface morphology. To efficiently search many atomistic
configurations on large length scales, we developed a parameter-free machine learning
tool for exploring organic/inorganic interfaces. Our preferred Bayesian optimisation approach [1] relies on probabilities to construct model functions, which are then iteratively
refined by input of real data points balancing exploitation with exploration. A Bayesian
optimisation algorithm was interfaced with both classical potential and density-functional
300
theory codes to enable iterative learning (on-the-fly)
of potential energy surfaces (PES)
on
3.6
large supercomputers without
human
input.
For
additional
efficiency,
the
method
exploits
250
3.5
d11
d13
structural
rigidity of molecular groups, reducing
the
degrees
of
freedom
and
making
the
200
3.4
d7
learning process analogous to “molecular LEGO“.
150
3.3
3.2
100
We present a proof-of-concept test based on the
alanine molecule (Figure 1) and an ap3.1
50 on the (101) surface of TiO
plication featuring electron donor C60 molecules
2 anatase.
The Bayesian optimisation d4
structure search (BOSS)
acquires PES information fast3 and is
0
2.9
particularly efficient in pinpointing the global minimum
structure. This versatile scheme
-50
0
100
200
300
for global minimum search could be extended beyond interface packing considerations to
address complex configurational problems across scientific disciplines.
E [kcal/mol]
a)
b)
300
3.6
250
d4 [deg]
d13
d4
3.5
200
3.4
150
3.3
100
3.2
50
3.1
3
0
2.9
-50
0
100
200
300
d13 [deg]
Figure 1: a) Illustration of two dihedral angles of the amino acid alanine; b) Corresponding potential energy map learned by BOSS after only 30 acquisition points.
[1] M. U. Gutmann and J. Corander, arXiv:1501.03291, stat.ML (2015)