using high-performance computing and molecular dynamics to rationalise micelle structure from small-angle scattering
SAS2018
— 2018/10/12
DOI:
10.6084/m9.figshare.7150823
Andrew R. McCluskey
Uni Bath/Diamond Light Source
arm61@bath.ac.uk
an_drewmcc
arm61.github.io/talks/sas2018-2
an_drewmcc
micellisation is tough to simulate
an_drewmcc
an_drewmcc
not a real potential energy surface
an_drewmcc
difficult to get realistic micelles without
apriori
information
easy to underestimate micelle size;
10.1016/j.theochem.2009.09.054
get
N
agg
from Guinier analysis;
10.1002/anie.201713303
run a few simulations at different
N
agg
pick the best agreement;
10.1002/anie.201713303
an_drewmcc
let's treat it like an optimisation problem
move the surfactants to optimise agreement with scattering profile
micelles are an ensemble structure, perfect for a population approach
population based optimisation algorithms are trivially parallelised
an_drewmcc
what are we optimising?
an_drewmcc
ctab
single tailed surfactant nearly has
C
∞v
symmetry
can describe the system with 5 numbers per molecule
three positional, two angular
an_drewmcc
particle swarm algorithm
an_drewmcc
each particle has knowledge of the personal best (within that member of the population) position and the global best position
apparently good for high dimensionality problems
an_drewmcc
shamelessly stolen from wikipedia
By
Ephramac
-
Own work
,
CC BY-SA 4.0
,
Link
an_drewmcc
fitoog
an_drewmcc
fitoog
fitoog is an open-source C_MPI code for applying a particle swarm algorithm fitting SAS data to atomistic/coarse-grained molecular models
runs a short energy minimisation using MARTINI force field parameters at each stage
github.com/arm61/fitoog
an_drewmcc
strong scaling
each node has 20 cores
an_drewmcc
weak scaling
each node has 20 cores
an_drewmcc
testing
an_drewmcc
is it good?
use fitoog to fit some "test data"
fifty CTAB molecules in a box
background subtracted SANS data
ten independent calcs, running on 16 nodes, population size 10240, number of steps 4000
an_drewmcc
Random numbers
acceptance probability 0.11 %
an_drewmcc
Random numbers w/ EM
acceptance probability 0.18 %
an_drewmcc
Particle swarm w/ EM
acceptance probability 0.26 %
an_drewmcc
not very micellar
an_drewmcc
an_drewmcc
conclusions
fitoog looks like a very efficient, pretty pointless code
we will test other algorithms, genetic
maybe try and optimise the particle swarm
can be used to produce test data for ML algorithms
an_drewmcc
acknowledgements
James Grant
Karen J. Edler
Stephen C. Parker
Andrew J. Smith
Jonathan L. Rawle
SCARF cluster
an_drewmcc
Sadie, my dog