Data Science Applications for Astronomy

Week 14: Artificial Intelligence for Mathematical & Physical Sciences:

Future or Hype?

Recent AI-themed Events

• Penn State's AI Week

• NSF Workshop on the Future of AI+MPS

National Science Foundation Directorates:

• Mathematical & Physical Sciences (MPS)

• Computer and Information Science and Engineering (CISE)

• Biological Sciences

• Engineering

• Geosciences

• Social, Behavioral and Economic Sciences

• STEM Education

• Technology, Innovation and Partnerships

Divisions of NSF CISE:

• Computer and Network Systems (CISE/CNS)

• Computing and Communication Foundations (CISE/CCF)

• Information and Intelligent Systems (CISE/IIS)

• Office of Advanced Cyberinfrastructure (CISE/OAC)

Divisions of NSF MPS:

• Astronomical Sciences (MPS/AST)

• Physics (MPS/PHY)

• Chemistry (MPS/CHE)

• Materials Research (MPS/DMR)

• Mathematical Sciences (MPS/DMS)

• Office of Strategic Initiatives (MPS/OSI)

NSF Workshop on the Future of AI+MPS

Themes:

• Interdisciplinary Research: Challenges and Opportunities

• Interdisciplinary Research: Resources Needed

• Education, Training, and Workforce

• Responsible AI

AI in the Astronomical Sciences

• How AST is Advancing AI

• How AI is advancing AST

• Progress at Intersection of AI & AST

How AST is Advancing AI

• Rich, open datasets

• Well-curated scientific literature

• Physics-informed machine learning frameworks

• Multi-modal data integration

• Experimental design optimization

• Multi-agent AI systems

How AI is advancing AST

Cosmology:

• Deep learning for parameter inferences on large scale structure & Comsic Microwave Background

• Neural network surrogate models for expensive N-body simulations

• Specialized architectures (e.g., Graph Neural Networks) to work with structure of cosmological simulation data sets

Time-domain Astronomy:

• Neural networks for real-time classification & early detection of transit events (e.g., supernovae, graviational waves)

• Irregular sampling and heterogeneous noise characteristics of astronomical time-series data (?)

• Unsupervised techniques (e.g., clustering algorithsm) to identify new classes of transients

Exoplanet Research:

• Machine learning to distinguish planetary signals from stellar variability

• Efficient exploration of highly multimodal posterior distributions (e.g., microlensing events)

Stellar Astrophysics:

• Spectral analysis of surveys (SDSS) to extract physical parameters from low-resolution and/or low signal-to-noise spectra

• Automating pipelines for analyzing eclipsing binary stars to search for stellar-mass black holes

• Neural networks to analyze stellar oscilation data to characterize stellar interiors and evolution

Galactic Evolution:

• Deep learning for extracting galaxy morphology from surveys

• Generating synthetic datasets for augmenting training data

• Reducing cost of inference with complex models

Heliophysics:

• Ensemble methods for Forecasting space weather

• Identifying and classifying sunspot groups and photospheric vector magnetic fields

• Neural fields for reconstruction of solar surface features to provide more complete picture of dynamics

Gravitational Waves:

• Deep learning for signal processing to separate true GW signals from noise

• Rapid identification and localization to support follow-up observations

• Reduce computational cost compared to matched filtering

Progress at Intersection of AI & AST

• Data processing scalability

• Anomaly detection & novel discovery

• Parameter inference through Simulation-based Inference

• Acceleration of simulations

• AI in control systems for instruments

Cross-Disciplinary Opportunities:

• Pursue the Science of AI

• Leverage AI for Conducting Research

• Promote Responsible AI

• Establish Robust AI Infrastructures

• Facilitate Interdisciplinary Collaborations

• Advocate for Diverse Funding Streams

• Educate and Train an AI+MPS Workforce

Pursue the Science of AI

• AI Innovations from Science

• Understanding AI

• Robust and Interpretable AI

Leverage AI for Conducting Research

• Hypothesis Generation

• Self-Driving Labs

• Synthesis and Communication

Promote Responsible AI

• Scientific Integrity

• Cost-Efficient Computing

• Public Engagement on AI+Science

Establish Robust AI Infrastructures

• Computing Resources

• Data Management and Access

• Benchmarking and Reproducibility

• Design Methods Optimized for AI for Science

  • Simulation-Based Inference

  • Uncertainty Quantification

  • Foundation Models

  • Reinforcement Learning for Experimental Control

  • AI for Quantum 2.0

  • Data-Efficient Methods

Facilitate Interdisciplinary Collaborations

• Research Opportunities

• Workshops and Conferences

• Knowledge Transfer

• Collaborating Beyond MPS

Advocate for Diverse Funding Streams

• Institute-Scale Activities

• Project-Scale Activities

• Individual Investigators

• Industry Collaborations

Educate and Train an AI+MPS Workforce

• Faculty Training

• Postdoctoral Training

• Graduate Education

• Undergraduate Education

• K-12 and Public Education

• AI Literacy

Opportunities in AI+AST

• Black-box nature of many existing methods

  • Developing more powerful and more efficient glass-box models that improve interpretability and explanability

  • Establishing metrics for evaluating AI

• Developing methods that are data-efficient

• Accuracy of AI is fundamentally limited by fidelity of simulations

  • If AI extracts more information from data, systematics likely more important

• Earning trust requires extensive testing and comparisons

  • Reliable uncertainty & bias quantification to understand limitations and robustness of AI approaches

  • Account for risk of bias in model development

  • Develop systematic frameworks for blind analyses.

• Creating systmes to actively participate in experimental design and potentially hypothesis generation

Setup