Do Androids Dream of Exploding Stars?

University of Delaware

Department of Physics and Astronomy

Biden School of Public Policy and Administration

Data Science Institute

LSST Survey Scientist

federica b. bianco

she/her

Grad student

Since 2019 we study the sky (and more!) mostly with AI

Postdoc

experiment driven science -∞:1900

theory driven science 1900-1950

data driven science 1990-2010

the fourth paradigm - Jim Gray, 2009

computationally driven science 1950-1990

experiment driven science -∞:1900

theory driven science 1900-1950

data driven science 1990-2010

the fourth paradigm - Jim Gray, 2009

computationally driven science 1950-1990

AI driven science? 2010...

Frank Rosenblatt, 1958

The Navy revealed the embryo of an electronic computer today that it expects will be able to walk, talk, see, write, reproduce itself and be conscious of its existence.

The embryo - the Weather Buerau's $2,000,000 "704" computer - learned to differentiate between left and right after 50 attempts in the Navy demonstration

NEW NAVY DEVICE LEARNS BY DOING; Psychologist Shows Embryo of Computer Designed to Read and Grow Wiser

July 8, 1958

what year did the first Neural Network in astronomy review came out?

Join at

slido.com

# 1366553

AI for survey design

AI for classification

1988

AI for classification

1988

CSP: Constraint Satisfaction Problems

1988

number of arXiv:astro-ph submissions with abstracts containing one or more of the strings: ‘machine learning’, ‘ML’, ‘artificial intelligence’, ‘AI’, ‘deep learning’ or ‘neural network’.

Smith+Geach May 2022 Astronomia ex machina

"In 1994, Ofer Lahav, an early trailblazer, wryly identified the ‘neuro-skeptics’—those resistant to the use of such techniques in serious astrophysics research—and argued that ANNs ‘should be viewed as a general statistical framework, rather than as an estoteric approach’ [8]. Unfortunately, this scepticism has persisted. This is despite the recent upsurge in the use of neural networks (and machine learning in general) in the field [...] Most of the criticism of machine learning techniques, and deep learning in particular, is levelled at the perceived ‘black box’ nature of the methodology."

physics

Input (observables)

Output

(observable)

Input (observables)

Output

(observable)

Data

Prediction

function

f(x)

Machine Learning

physics

f(x | M)

loss function (e.g. χ2, MSE)

Data

Prediction

f(x| m, b) = mx + b

m: slope

b: intercept

Machine Learning

f(x | M)

Data

Prediction

m: slope

b: intercept

parameters

learn

goal: find the right m and b that turn x into y

Machine Learning

f(x| m, b) = mx + b

f(x | M)

Data

Prediction

https://symposia.obs.carnegiescience.edu/series/symposium2/ms/freedman.ps.gz

"Regression toward Mediocrity":

Sir Francis Galton in the 1885-6

"Linear Regression":

Karl Pearson 1908-1911

Input

output

physics

f(x)

Data

Prediction

Input

output

physics

non-linear modification to a linear function

f(x) = \sigma( mx + b)

f(x)

Data

Prediction

Tree models

(at the basis of Random Forest

Gradient Boosted Trees)

Machine Learning

Galaxy Zoo

p(class)

extracted

features vector

p(class)

pixel values tensor

f(x)

GPT-3

175 Billion Parameters

3,640 PetaFLOPs days

Kaplan+ 2020

comput optimal frontier

Kaplan+ 2020

Input

output

hidden layers

latent space

A Neural Network is a kind of function that maps input to output accurately

Input

output

hidden layers

latent space

A Neural Network is a kind of function that maps input to output accurately

by generating "interesting" internal representation of the input

Input

output

hidden layers

latent space

loss function (e.g. χ2, MSE)

This latent space is only going to be as good as it needs to solve a particular task

Sam Altman:

(paraphrasing NPR interview 2025):

Well, what we learned was that we needed to get the model coherent and capable first. The content moderation, the safety stuff—that came after.

https://stratechery.com/2025/an-interview-with-openai-ceo-sam-altman-about-devday-and-the-ai-buildout/

What are they optimizing for?

back to astro......

trained extensively on large amounts of data to solve generic problems

Foundational AI models

We use the ILSVRC-2012 ImageNet dataset with 1k classes
and 1.3M images, its superset ImageNet-21k with
21k classes and 14M images and JFT with 18k classes and
303M high-resolution images.

Typically, we pre-train ViT on large datasets, and fine-tune to (smaller) downstream tasks. For
this, we remove the pre-trained prediction head and attach a zero-initialized D × K feedforward
layer, where K is the number of downstream classe

why not images too?

lightcurve latent space rep

image

latent space rep

SN 2018cow

Perley+2018

SN 2018cow

Perley+2018

late layers learn complex aggregate specialized features

early layers learn simple generalized features (like lines for CNN)

prediction "head"

original data

trained extensively on large amounts of data to solve generic problems

Foundational AI models

Adrian Crawford

UVA Jefferson fellow, PEO scholar

Beyond classification: Can Maven truly answer questoins it was not designed to answer?

can it tell us if a supernova is double peaked... even if we did not sample the peak?

FASTlab Flash highlight

Testing the performance of MetaAI SAM on astronomical objects

Instead of building our own specialized AI, can we adapt the models that the industry produces?

That would save a lot of computational resources and computational resources have an environmental footprint!

Award #2123264

ADS astronomy articles that mention "foundation model" in the abstract

LSST:

The Vera C. Rubin Observatory Legacy Survey of Space and Time

20Tb of data every night. That is equivalent to

8,000 high definition movies

4,000 hours of tiktok videos

every night for 10 years

what's in a name?

The first ground-based national US observatory named after a woman, Dr. Vera C. Rubin

what's in a name?

what's in a name?

The first ground-based national US observatory named after a woman, Dr. Vera C. Rubin

“The most important feature of any telescope is the imagination with which it is used.”

Building an unprecedented catalog of Solar System Objects

LSST Science Drivers

Mapping the Milky Way and Local Volume

Probing Dark Energy and Dark Matter

Exploring the Transient Optical Sky

To accomplish this, we need:

Objective: to provide a science-ready dataset to transform the 4 key science area

To accomplish this, we need:

1) Dark skies - Cerro Pachon Chile

Objective: to provide a science-ready dataset to transform the 4 key science area

To accomplish this, we need:

1) Dark skies - Cerro Pachon Chile

2) a large telescope mirror to be sensitive - 8m (6.7m)

Objective: to provide a science-ready dataset to transform the 4 key science area

May 2022 - Telescope Mount Assembly

3.2 Gigapixels:

We built the largest (declassified) camera ever built

to look farther and wider into the sky than ever before

In 2026 we will begin observing the sky with 1000 images every night for 10 years

1996-1998 Tony Tyson, Roger Angel

How it started

with Zhoran Mandami,

Astronaut Reid Wiseman,

Activist Zabib Musa Loro,

Pope Leo XIV,

Olympic medalist Alysa Liu,

Benicio Del Toro.......

2008

2017

Are We There YET????!!!!

Eye to the sky…on-sky engineering tests have begun at @nsfgov–@energy Rubin Observatory using the world’s largest digital camera!🔭

April 17 2025

Eye to the sky…on-sky engineering tests have begun at

Rubin Observatory using the world’s largest digital camera!

June 23 2025

First Look party here at UD with 213 people signed up!

678 separate images taken in just over seven hours of observing time. Trifid nebula (top right) and the Lagoon nebula, which are several thousand light-years away from Earth. | NSF-DOE Vera C. Rubin Observatory

June 30, 2025

DP1 release!

The Vera C. Rubin Observatory Data Preview 1

https://arxiv.org/pdf/2603.23786

https://arxiv.org/pdf/2507.22864v2

HELL YEAH!

2025

edge computing

Will we get more data???

SKA

(2025)

edge computing

Rubin LSST Transients by the numbers

17B stars (x10) Ivezic+19

~10 million QSO (x10) Mary Loli+21

~50k Tidal Disruption Events (from ~150) Brickman+ 2020

~10k SuperLuminous Supernovae (from ~200)Villar+ 2018

~400 strongly lensed SN Ia (from 10) Ardense+24

~50 kilonovae (from 2) Setzer+19, Andreoni+19 (+ ToO)

> 10 Interstellar Objects fom 2.... ?)

edge computing

Rubin LSST Transients by the numbers

17B stars (x10) Ivezic+19

~10 million QSO (x10) Mary Loli+21

~50k Tidal Disruption Events (from ~150) Brickman+ 2020

~10k SuperLuminous Supernovae (from ~200)

~400 strongly lensed SN Ia (from 10) Ardense+24

~50 kilonovae (from 2) Setzer+19, Andreoni+19 (+ ToO)

> 10 Interstellar Objects fom 2.... ?)

edge computing

Rubin LSST Transients by the numbers

17B stars (x10) Ivezic+19

~10 million QSO (x10) Mary Loli+21

~50k Tidal Disruption Events (from ~150) Brickman+ 2020

~10k SuperLuminous Supernovae (from ~200)

~400 strongly lensed SN Ia (from 10) Ardense+24

~50 kilonovae (from 2) Setzer+19, Andreoni+19 (+ ToO)

> 10 Interstellar Objects fom 2.... ?)

Rubin LSST Transients by the numbers

17B stars (x10) Ivezic+19

~10 million QSO (x10) Mary Loli+21

~50k Tidal Disruption Events (from ~150) Brickman+ 2020

~10k SuperLuminous Supernovae (from ~200) Villar+ 2018

~400 strongly lensed SN Ia (from 10) Ardense+24

~50 kilonovae (from 2) Setzer+19, Andreoni+19 (+ ToO)

> 10 Interstellar Objects fom 2.... ?)

Rubin LSST Transients by the numbers

17B stars (x10) Ivezic+19

~10 million QSO (x10) Mary Loli+21

~50k Tidal Disruption Events (from ~150) Brickman+ 2020

~10k SuperLuminous Supernovae (from ~200) Villar+ 2018

~400 strongly lensed SN Ia (from 10) Ardense+24

~50 kilonovae (from 2) Setzer+19, Andreoni+19 (+ ToO)

> 10 Interstellar Objects fom 2.... ?)

Rubin LSST Transients by the numbers

17B stars (x10) Ivezic+19

~10 million QSO (x10) Mary Loli+21

~50k Tidal Disruption Events (from ~150) Brickman+ 2020

~10k SuperLuminous Supernovae (from ~200) Villar+ 2018

~400 strongly lensed SN Ia (from 10) Ardense+24

~50 kilonovae (from 2) Setzer+19, Andreoni+19 (+ ToO)

> 10 Interstellar Objects fom 2.... ?)

SKA

(2025)

17B stars (x10) Ivezic+19

~10 million QSO (x10) Mary Loli+21

~50k Tidal Disruption Events (from ~150) Brickman+ 2020

~10k SuperLuminous Supernovae (from ~200) Villar+ 2018

~400 strongly lensed SN Ia (from 10) Ardense+24

~50 kilonovae (from 2) Setzer+19, Andreoni+19 (+ ToO)

> 10 Interstellar Objects fom 2.... ?)

True Novelties!

Rubin LSST Transients by the numbers

"BUT BIG DATA DOES NOT MEAN BIG SCIENCE"

Yang Huang,
University of Chinese Academy of Sciences

SpecCLIP talk @UNIVERSAI

IAU workshop Greece June 2025

survey optimization

Challenge

Introducing Rolling Cadence

Current plan: rolling 8 out of the 10 years

Discovery Engine

10M alerts/night

Community Brokers

target observation managers

Pitt-Google

Broker

BABAMUL

federica bianco - fbianco@udel.edu

the new astronomy discovery chain

https://www.lsst.org/content/charge-survey-cadence-optimization-committee-scoc

Rubin has involved the community to an unprecedented level in survey design this is a uniquely "democratic" process!

Survey Cadence Optimization Committee

Rubin LSST survey design

←Dimmer Brighter →

0.01 0.1 1 10 100

Rubin LSST survey design

2017

80,000

Rubin LSST survey design

Rubin has involved the community to an unprecedented level in survey design this is a uniquely "democratic" process!

2019

80,000

Rubin LSST survey design

Rubin has involved the community to an unprecedented level in survey design this is a uniquely "democratic" process!

2023

80,000

Rubin LSST survey design

Rubin has involved the community to an unprecedented level in survey design this is a uniquely "democratic" process!

2024

80,000

Rubin LSST survey design

Rubin has involved the community to an unprecedented level in survey design this is a uniquely "democratic" process!

2024

80,000

https://lssttooworkshop.github.io/

Rubin ToO program

https://lssttooworkshop.github.io/

Rubin ToO program

subverting the time domain astronomy process

Challenge

10 stars explode in the universe every second

Until the 1900s we would see 1 in a century

Until the 1980s we would see 1 in a decade

Until the 2010s we would see 1 in a month

With the Vera C. Rubin Observatory we will see 1000 every night !

in 60 seconds:

Difference Image Analysis

template

in 60 seconds:

Difference Image Analysis

template

difference image

in 60 seconds:

Difference Image Analysis

template

difference image

Rubin Observatory Data Management Team

federica bianco - fbianco@udel.edu

Improving the efficiency of transient detections with Neural Networks

template

difference

Saliency maps: what pixels matter?

template

difference

95% accurate

Acero-Cuellar et al. DESC submitted

Tatiana Acero-Cuellar

UNIDEL fellow,

LSST Data Science Fellow

FASTlab Flash highlight

Saliency maps: what pixels matter?

template

difference

Acero-Cuellar et al. DESC submitted

Tatiana Acero-Cuellar

UNIDEL fellow,

LSST Data Science Fellow

The Rubin LSST ML-Reliability Score (aka real-bogus)

accuracy 98.06%, purity 97.87%, completeness of 98.27%.

- requires instantaneous inference

- limited computational resources (CPU)

- evolving data quality

- limited ground truth data (e.g. no variable stars in training)

FASTlab Flash highlight

Challange

data encoding

well... it depends

2025

(2026)

edge computing

Is the data gonna also be better?

https://www.theverge.com/21298762/face-depixelizer-ai-machine-learning-tool-pulse-stylegan-obama-bias

data reconstruction with Generative AI

https://www.theverge.com/21298762/face-depixelizer-ai-machine-learning-tool-pulse-stylegan-obama-bias

data reconstruction with Generative AI

AI-assisted superresolution cosmological simulations

Yin Li+2021

LOW RES SIM

HIGH RES SIM

AI-AIDED HIGH RES

visualizatoin and concept credit: Alex Razim

Kaicheng Zhang et al 2016 ApJ 820 67

SN 2011fe

deSoto+2024

Boone 2017

7% of LSST data

Boone 2017

7% of LSST data

The rest

Kaggle PLAsTiCC challenge

AVOCADO classifier

https://arxiv.org/abs/1907.04690

Lochner et al 2018

https://arxiv.org/pdf/1812.00515.pdf

Kaggle PLAsTiCC challenge

AVOCADO classifier

https://arxiv.org/abs/1907.04690

Lochner et al 2018

https://arxiv.org/pdf/1812.00515.pdf

Text

Dr. Somayeh HKhakpash

LSST Catalyst Fellow

Lehigh University Visiting Prof.

FASTlab Flash highlight

Text

https://arxiv.org/pdf/2405.01672

we introduce

Gaussian process Optimized Photometric Regression of Extragalactic Archival Ultraviolet-infrared eXplosions, a.k.a GOPREAUX—

a Python package for Gaussian Process Regression of multi-wavelength transient photometry. [...]

This allows for predictions of light curves [...] at higher redshifts, where the rest-frame UV emission is redshifted into the observer-frame optical or infrared.

Gaussian processes work by imposing a kernel that represents the covariance in the data (how data depend on time or time/wavelength). Imposing the same kernel for different time-domain phenomena is principally incorrect

=> bias toward known classes

Methodological issues with these approaches

Neural processes replace the imposed kernel with a learned one

NASA FINESST Fellow

Siddharth Chaini

FASTlab Flash highlight

NASA FINESST Fellow

Siddharth Chaini

FASTlab Flash highlight

Challange

follow up

federica bianco - fbianco@udel.edu

Rubin will see ~1000 SN every night!

Credit: Alex Gagliano IAIFI fellow MIT/CfA

When they go high, we go low... spectra classification at low resolution

Astrophysical spectra require the capture of enough photons at each wavelength:

large telescopes

long exposure times

bright objects

Willow Fox Fortino

UDelaware

When they go high, we go low

Classification power vs spectral resolution for SNe subtypes

FASTlab Flash highlight

Willow Fox Fortino

UDelaware

When they go high, we go low

Classification power vs spectral resolution for SNe subtypes

Willow Fox Fortino

UDelaware

When they go high, we go low

Classification power vs spectral resolution for SNe subtypes

Adapting Transformer architecture (Vaswani et al. 2017)

Kaggle PLAsTiCC challenge

AVOCADO classifier

https://arxiv.org/abs/1907.04690

Classification from sparse data: Lightcurves

Viswani+ 2017 Attention is all you need

AI was transformed in 2017 by this paper

Willow Fox Fortino

UDelaware

When they go high, we go low

Classification power vs spectral resolution for SNe subtypes

FASTlab Flash highlight

Willow Fox Fortino

UDelaware

When they go high, we go low

Classification power vs spectral resolution for SNe subtypes

FASTlab Flash highlight

Willow Fox Fortino

UDelaware

When they go high, we go low

Classification power vs spectral resolution for SNe subtypes

data embedding

classification head

Willow Fox Fortino

UDelaware

As seen in Muthukrishna+2019

FASTlab Flash highlight

Text

A new AI-based classifier for SN spectra at low resolution

we badly need better benchmark datasets

anomaly detection

Challenge

Most classifiers for variable stars use Random Forest (not distance based)

In distance based classification, optimal distances can be found for the class of interest: flexible, customizable, efficient

https://arxiv.org/pdf/2403.12120.pdf

Astronomy and computing

FASTlab Flash highlight

NASA FINESST Fellow

Siddharth Chaini

UDelaware

Text

Are we prepared to discover new physics?

Text

Are we prepared to discover new physics?

This ensamble distance method excells at identifying out of sample anomalies!

Text

Are we prepared to discover new physics?

Chaini, Bianco, and Mahabal

submitted to MLPS NEURIPS 2025

NSF award

2219731

NASA FINESST Fellow

Siddharth Chaini

UDelaware

This ensamble distance method excells at identifying out of sample anomalies!

Text

Are we prepared to discover new physics?

Chaini, Bianco, Mahabal

submitted to MLPS NEURIPS 2025

https://www.epa.gov/energy/greenhouse-gas-equivalencies-calculator#results

https://arxiv.org/pdf/2602.10181

https://arxiv.org/abs/2601.02484

Will AI replace us (astrophysicists?)

is a word I am borrowing from Margaret Atwood to describe the fact that the future is us.

However loathsome or loving we are, so will we be.

Whereas utopias are the stuff of dream dystopias are the stuff of nightmares, ustopias are what we create together when we are wide awake

https://www.youtube.com/watch?v=QO3nY_u6hos

US-TOPIA

thank you!

University of Delaware

Department of Physics and Astronomy

Biden School of Public Policy and Administration

Data Science Institute

federica bianco

fbianco@udel.edu

https://slides.com/federicabianco/uvacolloquium2026

Opportunity

foundational models

ISN'T FREE

ethics of AI

Challange + Opportunity

Knowledge is power

Astrophysical data is a sandbox. It has no social value, no privacy risk. We can safely learn about how bias builds into algorithm and how to correct it

Knowledge is power

Astrophysical data is a sandbox. It has no social value, no privacy risk. We can safely learn about how bias builds into algorithm and how to correct it
Ethics of AI is a critical element of the education of a technologist

With great power comes grteat responsibility

"Sharing is caring"

Astrophysical data is a sandbox. It has no social value, no privacy risk. We can safely learn about how bias builds into algorithm and how to correct it
Ethics of AI is a critical element of the education of a technologist
AI is a transferable skill - use if for good!

the butterfly effect

We use astrophyiscs as a neutral and safe sandbox to learn how to develop and apply powerful tool.

Deploying these tools in the real worlds can do harm.

Ethics of AI is essential training that all data scientists shoudl receive.

https://www.theverge.com/21298762/face-depixelizer-ai-machine-learning-tool-pulse-stylegan-obama-bias

Why does this AI model whitens Obama face?

Simple answer: the data is biased. The algorithm is fed more images of white people

But really, would the opposite have been acceptable? The bias is in society

models are neutral, the bias is in the data (or is it?)

https://www.theverge.com/21298762/face-depixelizer-ai-machine-learning-tool-pulse-stylegan-obama-bias

Why does this AI model whitens Obama face?

Simple answer: the data is biased. The algorithm is fed more images of white people

But really, would the opposite have been acceptable? The bias is in society

models are neutral, the bias is in the data (or is it?)

https://www.theverge.com/21298762/face-depixelizer-ai-machine-learning-tool-pulse-stylegan-obama-bias

Why does this AI model whitens Obama face?

Simple answer: the data is biased. The algorithm is fed more images of white people

Joy Boulamwini

models are neutral, the bias is in the data (or is it?)

Challange

echological AI

←Dimmer Brighter →

0.01 0.1 1 10 100

stellar sexplosions

stellar eruptions

stellar variability

trained extensively on large amounts of data to solve generic problems

Foundational AI models

We use the ILSVRC-2012 ImageNet dataset with 1k classes
and 1.3M images, its superset ImageNet-21k with
21k classes and 14M images and JFT with 18k classes and
303M high-resolution images.

Challenges in Space-Based Observations

Limited Field of View: Space telescopes often have smaller fields of view compared to ground-based surveys.
Data Latency: Delays in data transmission and processing can affect rapid follow-up.
Resource Allocation: Competition for telescope time can limit observations of certain transients.... LETS NOT TRIGGER 3 ToOs ON THE SAME TRANSIENT!!