•S uc u ed me ada a o desc ibe emula o iles.
•Also o ms con igu a ion ile used o inpu o emula o aining and
scien i ic in e ence amewo ks.
•Desc ibed in publica ion. doi:10.1093/ as i/ za 002
•De ailed in public gi hub. h ps://gi hub.com/HTJense/cosmopowe
•Ve sioned elease in p ep.
•In p inciple can p opaga e his emula o s anda d ou side o as ophysics!
Typical ques ions o a po en ial use :
•Wha emula o s e en exis ?
•Wha code and model is his emula o emula ing?
•Which e sion?
•Wha ange o pa ame e alues is i alid o ?
•Which obse ables does i ou pu ?
•Wha happens i I y and use i ou side hese speci ica ions?
•Can I ex end i ?
Typical consequences:
•Failed MCMC uns (was ed esou ces).
•E oneous in e ence om mis-unde s anding
which model is being used.
•Con usion when alida ing esul s. Why does
my esul di e o you s?
•Was ed e o and esou ces on e-c ea ing
own (and hence us ed) e sion o emula o s.
Enhancing Open Cosmology wi h
Emula o Packaging
Ian Ha ison†, Hidde T. Jense
The Science: Cosmology wi h Angula Powe Spec a
S a e o he A : Compu a ional Speed-up wi h Emula o s
S a emen o Need: Reliable Emula o Sha ing and Re-Use
Solu ion: FAIR- ocussed Emula o Packaging
Implemen a ion: Lowe ing Ba ie s wi h an Open F amewo k
•Each majo science analysis ypically in ol es analysis o ~100 chains which ake
~1000 co e hou s each on HPC clus e s. The main compu a ion ime is om
nume ically sol ing he Eins ein-Bol zmann equa ions.
•T adi ional code op imisa ion o Eins ein-Bol zmann sol e s is eaching limi s due o
inc easing ins umen al p ecision equi ing be e accu acy and ewe app oxima ions.
•No el physics models in ol e adding new ing edien s (e.g. exo ic ma e )
o he Uni e se and a e e en slowe o sol e o .
•Inc easingly sol ed wi h ained Neu al Ne wo k
Emula o s, which lea n om
p e-gene a ed aining da a and ~ins an aneously p edic da a om model
.
•Valida ion o exis ing ne wo k plus me ada a.
•C ea ion o aining se s.
•T aining o ne wo ks.
•Plo ing o accu acy wi h espec o emula ed code.
•In e aces o common in e ence codes.
•Imp o e pedagogy h ough making in e ence possible on a
lap op!
•Quickly es new da a and e ec s o al e na i e
modelling choices.
•Allow o he s o use new and no el physical models.
•C ea ing and p opaga ing an emula o o you no el model is
much easie han p opaga ing he al e ed nume ical code.
Da a
:
In ensi y ields
on he sphe e
(Cosmic
Mic owa e
Backg ound
empe a u e)
Model
:
E olu ion o he
Uni e se using
Eins ein-Bol zmann
Equa ions
Compa ison space
:
Angula powe spec a
Ou pu s
:
Con en s, his o y o he Uni e se
Communi y
:
Cosmologis s and as ophysicis s
ac oss mul iple ~100-pe son
in e na ional collabo a ions
Emula o s usually
c ea ed and sha ed on ad-hoc
pe son- o-pe son basis!
We implemen ed a use - iendly amewo k o he end- o-end
wo k low o NN emula o c ea ion, sha ing and use.
Model pa ame e s (e.g. ac ion o Uni e se made o da k ma e ) a e in e ed using
Ma ko Chain Mon e Ca lo pos e io es ima ion, compa ing models o powe spec a o da a.
We ha e c ea ed s anda dised “packaging” o
Neu al Ne wo k emula o s.
†ha isoni@ca di .ac.uk
School o Physics and As onomy, Ca di Uni e si y, The Pa ade, Ca di , Wales, UK CF24 3AA
P i . comm. 2023?
CosmoPowe
doi:10.1093/mn as/s ac064
Uni e ses wi h
di e en expansion
speeds NASA/WMAP Science Team
