Cosmology's Midlife Crisis
IAIFI Fellow, MIT
Carolina Cuesta-Lazaro
Art: "A Bar at the Folies-Bergère" by Édouard Manet
Embracing the Machine Learning Makeover
Robust
Very few assumptions about the mechanisms of galaxy formation
Generalizable
Physical Inductive biases
Symmetries: Homogeneous and isotropic Universe on large scales
Precise
(On large scales)
Information Loss
Galaxy formation = Nuisance
Due to being fast
The perks of Perturbation Theory
Small Scales
High dimensional inference
Carolina Cuesta-Lazaro IAIFI/MIT @ Split 2024
p(G∣δdm,A,z)
p(\mathcal{G}|\delta_\mathrm{dm}, \mathcal{A}, z)
Galaxy properties
Astro parametrization
DM density + velocity fields
Target
1. Robust: Mapping dark matter into anything
Carolina Cuesta-Lazaro IAIFI/MIT @ Split 2024
MTNG
Astrid
UniverseMachine
CAMELS
Data
Hydro Sims
SAMS
Santa Cruz
p(x0)
p(x_0)
p(x1)
p(x_1)
Dark Matter
"Baryons"
xt=tx0+(1−t)x1
x_t = t x_0 + (1-t) x_1
dtdxt=ut(xt)
\frac{d x_t}{dt} = u_t(x_t)
Transformer
G={x,v,SFR,M⋆}
\mathcal{G} = \{x, v, \mathrm{SFR}, M_\star \}
Carolina Cuesta-Lazaro IAIFI/MIT @ Split 2024
t
t
How many parameters can all these simulations share?
Carolina Cuesta-Lazaro IAIFI/MIT @ Split 2024
MTNG
Astrid
UniverseMachine
CAMELS
Data
Hydro Sims
SAMS
Santa Cruz
Can we fit the remaining to observations directly?
2. Generalizable: Representation learning
Learning to represent feedback
p(G∣C,A)
p(\mathcal{G}|\mathcal{C}, \mathcal{A})
Carolina Cuesta-Lazaro IAIFI/MIT @ Split 2024
SIMBA
TNG
Astrid
Different Supernovae and AGN feedback mechanisms
+ Different parametrisations
Finding anomalies
Z
pΛCDM(x)p(x)
\frac{p(x)}{p_{\Lambda\mathrm{CDM}}(x)}
x
x
"A point cloud approach to generative modeling for galaxy surveys at the field level"
Cuesta-Lazaro and Mishra-Sharma
arXiv:2311.17141
Base Distribution
Target Distribution
- Sample
- Evaluate
3. Physical Inductive biases
Carolina Cuesta-Lazaro IAIFI/MIT @ Split 2024
Same model architecture (transformer) on direct inference wouldn't work at all!
Carolina Cuesta-Lazaro IAIFI/MIT @ Split 2024
p(x∣C)
p(x|\mathcal{C})
μθ(zt,t)
\mu_\theta(z_t,t)
p(zT)
p(z_T)
Making homogeneous and isotropic Universes
Base Distribution
Denoiser
=
=
p(
p(
)
)
p(
p(
p(
p(
)
)
Invariant
Equivariant
=
=
p(
p(
)
)
p(
p(
=
=
p(
p(
)
)
p(
p(
Carolina Cuesta-Lazaro IAIFI/MIT @ Split 2024
All learnable functions
All learnable functions constrained by your data
All Equivariant functions
More data efficient!
Carolina Cuesta-Lazaro IAIFI/MIT @ Split 2024
Equivariant models
Non-equivariant
SEGNN: arXiv:2110.02905
NequIP: arXiv:2101.03164
Carolina Cuesta-Lazaro IAIFI/MIT @ Split 2024
Robust
Very few assumptions about the mechanisms of galaxy formation
Generalizable
Physical Inductive biases
Symmetries: Homogeneous and isotropic Universe on large scales
Precise
(On large scales)
Information Loss
Galaxy formation = Nuisance
Due to being fast
The perks of Machine Learning
Small Scales
High dimensional inference
Carolina Cuesta-Lazaro IAIFI/MIT @ Split 2024
Learned
Learned representations
Symmetries: Homogeneous and isotropic Universe on large scales
as long as we can sufficiently simulate
Interesting Physics
Perturbation Theory's retirement plan:
The Physics of Machine Learning
Carolina Cuesta-Lazaro IAIFI/MIT @ Split 2024
Grokking
Double Descent
Emergent abilities
Cosmology's Midlife Crisis IAIFI Fellow, MIT Carolina Cuesta-Lazaro Art: "A Bar at the Folies-Bergère" by Édouard Manet Embracing the Machine Learning Makeover
Split-Croatia2024
By carol cuesta
Split-Croatia2024
- 306
