Pop-cosmos: Machine-learning the galaxy population for large cosmological surveys
Dr Stephen Thorp (IoA and Kavli Institute for Cosmology, Cambridge)
Projects such as the Vera C. Rubin Observatory and Euclid are critical tools for understanding cosmological questions like the nature of dark energy. By observing huge numbers of galaxies, they enable us to map the large scale structure of the Universe. However, this is only possible if we are able to accurately model our photometric observations of the galaxies. In this talk I will present the "pop-cosmos" framework, a generative model for the galaxy population, which we have been developing to tackle this challenge. I will show how we can write down a flexible physics-based recipe for rapidly synthesizing large samples of mock galaxies, how we can "train" this model to match our observations, and what we can do with such a model once it is trained. Along the way, I will try to highlight some of the key techniques and technologies that make this work possible: neural networks that emulate computationally costly physical models; generative machine learning tools such as "diffusion" models; and graphics processing unit (GPU)-based computing.
Stephen Thorp is a postdoctoral researcher at the Institute of Astronomy and Kavli Institute for Cosmology, Cambridge, where he works with Prof. Hiranya Peiris. He spent the first three years of his postdoctoral position in the Oskar Klein Centre at Stockholm University, and relocated to Cambridge in September 2025. From 2018-2022 he was a PhD student at the Institute of Astronomy, Cambridge, working with Prof. Kaisey Mandel. Before that he studied physics at the University of Birmingham. Stephen has broad interests in statistics and machine learning applied to astronomy problems, particularly in the fields of galaxy evolution (which he'll be talking about at the RAS), strongly-lensed transients, and supernova cosmology.
From background seismicity to seismic crises: What can we learn from Machine Learning-Enhanced Catalogs?
Francesco Scotto di Uccio
Microseismicity continuously occurs along the seismogenic structures that can also host larger, destructive earthquakes. Therefore, its characterization can potentially provide crucial information on the geometry and mechanical state of the underlying faults, before the occurrence of notable events. However, conventional catalogs are limited in size, since many small events are hidden in the noise. Therefore, discovering such events calls for advances in both earthquake detection techniques and monitoring infrastructures.
We showcase how the integration of machine learning and similarity-based detection techniques can increase the content of seismic catalogs both for background seismicity and seismic sequences, up to one order of magnitude as compared to conventional manual catalogs. In this framework, machine learning models provide an enhanced set of events as compared to the standard catalog, which can be used as templates to effectively detect lower magnitude, colocated events, also with a short interevent time. The location of the earthquakes in the enhanced catalog revealed the activated fault patches, while the determination of source properties enabled the definition of evolutive models for the seismic sequences. To monitor background seismicity, we integrated the standard seismic network with 200 stations deployed in dense arrays for one year, demonstrating the possibility to consistently detect small magnitude earthquakes with the use of short-term dense arrays and established machine learning models. Moreover, this configuration enables the downscaling of the seismicity characteristics to small, decametric-size events, achieving resolution as from multiple years of conventional monitoring.
We finally focused on the characterization of the 2025 Santorini seismic sequence, revealing the intense level of seismic activity during the crisis and demonstrating how deep-learning approaches can support near–real-time tracking of the evolution and dynamics of the sequence.
I am currently a PostDoc Research at the Department of Physics of the University of Naples Federico II, where I obtained my Bachelor’s and Master’s degrees in Physics, followed by a PhD in Structural, Geotechnical, and Seismic Risk Engineering (awarded on February 21, 2025), with a thesis entitled “Detection and characterization of microseismicity using advanced techniques” (supervisors: Prof. Gaetano Festa and Prof. Matteo Picozzi). During my PhD program, I carried out research periods as a Visiting Student at Stanford University and at the GFZ German Research Centre for Geosciences in Potsdam, where I had the opportunity to develop and apply advanced techniques for microseismicity detection, also using fiber-optic sensors for seismic monitoring.
My scientific activity focuses on the characterization of small-magnitude earthquakes through the implementation of artificial intelligence techniques for the automatic identification of earthquakes within continuous seismic signals and the generation of high-density seismic catalogs. These catalogs enable accurate estimation of event hypocenters, assessment of the energy released during the rupture process, and analysis of the geometrical properties of seismogenic sources. The results of this research have led to the development of an automatic strategy for the comprehensive characterization of seismicity in the area affected by the destructive 1980 Irpinia earthquake (Southern Italy), which is currently monitored by a dense seismic network (Irpinia Near Fault Observatory), providing continuous support for monitoring activities. Moreover, I am actively involved in national research groups focused on the development of advanced strategies for the automatic identification of seismicity using machine learning models, as well as on the use of fiber-optic sensors for continuous monitoring.
Revealing Hidden Information in Volcanoes and Faults with Machine
Learning
Dr. Zahra Zali
Seismology is a strongly data-driven science, and the rapid growth
of seismic observations offers new opportunities to learn more from Earth
signals. Modern machine-learning methods enable faster and deeper analysis
of large and complex datasets than is often possible with classical
approaches. In this presentation, I introduce a deep-learning approach for
the automatic analysis of continuous seismic data, applicable across
different data types and tectonic settings. Using the 2021 Iceland eruption
as a volcanic case study, I show how this approach can identify key eruptive
phases and detect weak precursory tremor signals. I then present examples
from active faults, including the East Anatolian Fault Zone and the San
Andreas Fault, where the same approach reveals subtle patterns that are
difficult to detect with traditional techniques. Together, these examples
show how such
analyses provide new insight into transient processes and stress evolution
in the solid Earth.
Dr. Zahra Zali is a postdoctoral researcher in seismology at GFZ
Helmholtz Centre for Geosciences. Her research focuses on the analysis of
continuous seismic and strain data to study transient processes in faults
and volcanoes, with an emphasis on developing machine-learning–based
methods. She has worked on volcanic eruptions, slow deformation, and
earthquake-related signals across a range of tectonic settings. Her current
research aims to detect and characterize the preparatory processes preceding
earthquakes and volcanic eruptions.