Highlights from the AASKAII chapter

"The Role of Artificial Intelligence in the SKA Era"

|
2026-08-31 | Philipp Denzel

Content

Chapter (Methods and Techniques) in AASKAII
Advancing Astrophysics with the SKA – II
Overview of AI applications for the SKAO
Six scientific problem areas, each mapped onto AI-based remedies
From current problem to future directions

Message

"AI is not only an automation tool for coping with scale.
It is a catalyst for discovery."

Operational necessity — real-time triage, faint signals under extreme compression, calibration under time-variable systematics

Scientific leverage — non-Gaussian information, accelerated forward modelling, uncertainty-aware field-level inference

Imaging & sky simulations — CLEAN scales poorly, needs manual convergence, no calibrated confidence
Epoch of Reionization / 21-cm — foregrounds 4–5 orders brighter; robustness bought with sensitivity
Radio galaxies — PyBDSF/Aegean assume fixed morphology, break on multi-component and rare sources
Pulsars & transients — RFI, rare-event imbalance, real-time latency
Cosmic magnetism — faint Stokes V and RM signals under instrumental artifacts
Gravitational lensing — thousands of radio lenses expected; radio AI work still scarce vs. optical

AI for mitigating the data deluge

A global federated computing fabric planned a decade before first light — WLCG: ~1.4 M cores, 1.5 EB, 170+ sites. SRCNet is the analogue.
A shared software stack and a real open-data and reproducibility culture

Lessons not yet in focus

CERN retrofitted ML into the system design
SKAO can be AI-native at design time — learned triage in the SDP, specified now, not 2040

The LHC trigger keeps roughly 1 event in 20 000. Everything else is discarded, permanently, by cuts hand-designed in the 2000s.

But: an AI trigger that discards data must be uncertainty-aware and auditable, or the "unknown unknowns" go first

1st

SKA+AI papers per capita

SKACH institutions

2nd

SKA+AI papers per institution with A&A programme

Member since Jan 2022 · ~100 people · contributions concentrated in science with AI, the software and data layer (Karabo, SDP, SRCNet)

AI is a prerequisite for SKA-scale science, not an optimization.
Unlike CERN, we get to design for it from the start — so the burden shifts to explainability, uncertainty quantification and MLOps, now!
Switzerland concentrates talent for AI and the software layer.