Galaxy clusters are the most massive gravitationally bound structures in the Universe and are dominated by dark matter, which accounts for nearly 90% of their total mass. When sufficiently massive, galaxy clusters become powerful strong gravitational lenses, producing multiple images and highly magnification of distant background galaxies. These strong lensing clusters provide measurements of the dark matter distribution on megaparsec scales, offering unique insights into the formation and evolution of cosmic structure. Strong lensing clusters also act as natural telescopes, magnifying galaxies in the early Universe and even individual stars that would otherwise be too faint to observe. Beyond their astrophysical applications, strong lensing clusters have become increasingly powerful cosmological probes, allowing us to study both the geometry of the Universe and its expansion history through multiply imaged galaxies and lensed transients.

One particularly exciting application is cluster strong lensing cosmography. Because background galaxies span a wide range of redshifts, the positions of their multiple images depend on the cosmological distance. Precise lens models therefore enable direct constraints on cosmological parameters that govern the geometry of the Universe, including the total matter density, Ωm, and the properties of dark energy. As illustrated in Figure 1 (Caminha et al., 2022), this technique provides an independent and complementary probe to Type Ia supernovae, baryon acoustic oscillations, and the cosmic microwave background, with strong potential to improve our understanding of the fundamental cosmological model.

Our group is actively developing precision strong lensing models of galaxy clusters for time-delay cosmography. In particular, we have constructed a detailed mass model of the galaxy cluster MACS J0138.0−2155 (Ertl et al. 2025), which hosts the remarkable strongly lensed supernovae SN Requiem and SN Encore in the same background galaxy. By combining high-resolution imaging, spectroscopy, and advanced lens modeling technique, we accurately predict the time delays and magnifications of the multiple supernova images. Together with measured time delays, these models enable a precise determination of the Hubble constant, H₀, providing a new and independent approach to addressing the current Hubble tension (Suyu et al. 2026).

Understanding the mass assembly history of galaxies from the distant to the local Universe is a major research axis in extragalactic astrophysics. Observational and theoretical studies have now established a consistent evolutionary sequence for galaxies in the upper-end of the mass distribution. Massive galaxies have experienced a rapid growth phase during the first 2-3 Gyr after the Big Bang, with major mergers and extraordinary amounts of dust and gas triggering extreme star formation episodes (de Lucia et al. 2006). When the gas reservoirs got depleted, this intense phase of compact, nuclear star formation ceased, leaving red quiescent galaxies (Toft et al. 2007). Their subsequent evolution was driven by minor mergers, mainly with low-mass and gas-poor galaxies, which induced substantial growth in size with moderate increase in stellar mass (Naab et al. 2009). After 8-10 Gyrs, this phase resulted in the massive elliptical galaxies seen in the local Universe.

We are using multi-wavelength observing programs and detailed modeling to address fundamental questions about the physics underlying this evolutionary sequence. In particular, we are using the strong lensing effect to probe: (1) the populations and properties of high-redshift lensed galaxies, and (2) foreground galaxies acting as light deflectors.

First of all, with the boost in sensitivity and angular resolution from strong lensing, we obtain fine diagnostics on the internal properties of galaxies during their initial growth phase. With observations from the optical/near-infrared to the millimeter and radio, we are studying the different components of the interstellar medium (atomic, molecular, and ionized gas, dust content) as well as the stellar populations. We have for instance obtained advanced constraints on the molecular gas phases in these galaxies, including for the dense cloud cores (Cañameras et al. 2018a, 2021), and studied the star formation law down to the small scales of ~100 pc (Figure 1; Cañameras et al. 2017b). By combining these data sets with detailed modeling, we also probe the feedback mechanisms from star formation and/or active galactic nuclei. Recent advances in lensed galaxy reconstruction techniques are also expected to reveal more of their intrinsic morphology (Galan et al. 2021).

Secondly, the strong lensing effect offers a unique chance to obtain robust mass estimates for the foreground lens galaxies, and to gain insights into their baryonic and dark matter components. We are giving a particular focus on the most massive ellipticals residing in overdense environments. For instance, with HST imaging and MUSE spectroscopy, we have shown that joint modeling of strong lensing observables and spatially-resolved stellar kinematics with GLEE (Suyu et al. 2010, 2012) and GLaD softwares (Chirivi et al. 2020, Yildirim et al. 2020) allows us to firmly separate galaxy-scale and extended dark-matter halos (Figure 2; Wang et al. 2022). At the scale of individual galaxies, strong lensing enables the detection of invisible structures such as dark galaxy satellites and complex variations in the mass distribution of the lens (Galan et al. 2022). This opens new avenues to test the influence of the past merging history and the environment on the mass distribution of massive galaxies.

As part of our broader work on lensed supernovae, we search for transients in massive strong-lensing galaxy clusters. Compared with quasars, supernovae evolve rapidly and have well-characterized light curves, offering the potential for precise time-delay measurements when the supernovae are strongly lensed. Galaxy clusters are particularly promising targets because they can strongly magnify distant background galaxies and produce widely separated images with relatively long time delays. Their many member galaxies also increase the chance of detecting a Type Ia supernova occurring in a cluster member galaxy.  This provides an opportunity to measure the (luminosity) distance to the cluster via the Type Ia SN, in addition to the time-delay distance derived from a strongly lensed variable source that is lensed by the same cluster  Such systems with dual distance measurements are a powerful cosmographic probe (Acebron et al. 2023).

We currently monitor three strong-lensing clusters. SDSS J1029+2623 and COOL J0335−1927 are observed with the VLT Survey Telescope through a program led by Claudio Grillo. Both systems produce multiple images of a background quasar and have existing lens models and time-delay predictions (Figure 1; Acebron et al. 2022; Napier et al. 2023). We also monitor MACS J0138 with the MPG/ESO 2.2-metre telescope at ESO’s La Silla Observatory through a program led by our research group.

To discover new transients, we compare observations taken on different nights with a deep reference image. After matching the images in position, brightness scale, and point-spread function, we subtract the reference image so that static sources largely disappear, while supernovae and other variable objects remain visible. An example of this image-subtraction process is shown in Figure 2. We are developing an automated pipeline to generate these difference images, identify transient candidates, and extract their light curves.
 

Supernova progenitors

Even though Type Ia SNe are a powerful cosmological probe that led to the discovery of dark energy (Riess et al. 1998, Perlmutter et al. 1999), their progenitors are still a puzzle after decades of debate. Are they single-degenerate (SD) systems with a white dwarf (WD) accreting material from a non-degenerate companion and exploding when approaching the Chandrasekhar mass limit (e.g., Whelan & Iben 1973), or double-degenerate (DD) systems with two WDs merging (e.g., Tutukov & Yungelson 1981; Iben & Tutukov 1984), or other mechanisms?

Core-collapse progenitors, on the other hand, are well established to be massive stars with ZAMS > 8-10  M⊙. However, the final stages of their evolution are still poorly constrained. There has been growing evidence for intense activity in the final weeks to centuries prior to core collapse, either in the form of episodic outbursts or steady-state mass-loss via stellar winds (e.g., Smith 2014). The interaction of the SN ejecta with the resulting circumstellar material can dramatically impact the radiative display, especially during very early phases of the SN evolution.

Strongly lensed SNe enable observations of a SN explosion right from the beginning, which was nearly impossible to do in the past, given both the difficulty of finding SNe at very early phases and the time lag to arrange follow-up observations after a SN is detected.  By exploiting the time delay between the multiple SN images, a lensed SN system can be detected based on the first appearing SN image, and follow-up (especially spectroscopic) observations of the next appearing SN image can be obtained right from the explosion. Early-phase observations are crucial for understanding the progenitors of SNe and their mass-loss history.

One of the goals of the HOLISMOKES program (www.holismokes.org) and the main focus of the newly started SuperEarly program is to acquire such early-phase observations of lensed SNe to unveil their progenitors. Through comprehensive microlensing simulations of a suite of SN Ia explosion models, Suyu et al. (2020) showed that microlensing distortions on the early-phase spectra of SNe Ia would typically be negligible, within 1% (1-sigma spread). This provides excellent prospects for acquiring intrinsic early-phase SN Ia spectra through strong lensing to study the progenitors of SNe Ia.

When a star passes sufficiently close to a massive black hole, it can be captured or disrupted. If the star is disrupted outside of the black hole event horizon, known as a tidal disruption event (TDE), some of the stellar material is bound and produces a luminous flare as the stellar debris falls back onto the black hole. These astrophysical transients produce a characteristic light curve that remains observable on the order of a year.

TDEs were a purely theoretical speculation by the time their theoretical model was first proposed in the 1980’s (e.g., Lacy et al. 1982, Hills 1988, Rees 1988). Since then, the events have been detected at an accelerating rate; roughly 100 events are known so far. A number that will soon be outdated, as it will continue to grow at a much faster pace with detections from ongoing surveys (e.g., eROSITA, ZTF, and especially Rubin Observatory’s LSST). In particular, LSST is expected to increase the TDE sample size by a few orders of magnitude (Bricman & Gomboc 2020). Since some of the candidates could be lensed events, it is important to have reliable estimates for the detection rate of lensed TDEs. We have improved on current estimates of lensed TDE rates, showing that LSST will observe lensed TDEs (Szekerczes et al. 2024; Mamuzic et al. 2025). 

The uncertainties in this estimate largely stem from a theoretical unknown: the emission mechanism powering TDEs. The comparison between our estimates and the observed rate in the future will place strong constraints on the underlying mechanism. In addition, we consider potential microlensing events in strongly lensed TDEs to constrain the sizes of their emission regions. This will allow us to put additional, direct limitations on the emission mechanism in TDEs from a single event.

Machine learning is becoming an essential tool in astronomy, driven by the rapidly increasing volume and complexity of data from large astronomical surveys. Deep learning, in particular, enables the efficient analysis of massive data sets and the detection of subtle spatial, temporal, and spatio-temporal signals across multiband observations. We actively develop and apply these methods to strong gravitational lensing, from discovering static lenses and lensed transients, to measuring precise time delays between multiple images of lensed transients, and modelling strong-lensing systems.

For the discovery of galaxy- and cluster-scale strong lenses, we develop and apply state-of-the-art deep learning methods, including convolutional neural networks (CNNs), ResNet architectures, vision transformers, and variational autoencoders (Canameras et al. 2020, 2021, 2023; Shu et al. 2022, Schudlt et al. 2024, 2025). Our work has led to thousands of new high-quality strong-lens candidates in surveys including Pan-STARRS, HSC, and Euclid, with extensions to LSST ongoing. We are also exploring how to combine the high spatial resolution of Euclid with the multiband colour information from ground-based surveys such as LSST to improve the detection of strong lenses (Melo et al. 2025). Deep learning is also being used for automated strong-lens modelling, enabling efficient and scalable inference of lens properties for large samples (Gu et al 2022, Galan et al. 2022).

For the discovery of lensed transients, our OASIS (On-the-fly ASynchronous Image Stream classification) framework targets the spatio-temporal signatures of transients in asynchronously sampled, multiband image streams. We have explored recurrent architectures combining ConvLSTM and LSTM layers (Bag et al. 2025, 2026), as well as transformer-based models, the latter showing particularly promising performance and flexibility (Bag et al. 2026 in prep). The framework is designed for LSST and will serve as a backbone for direct searches for lensed transients in the LSST alert stream, where hundreds of lensed supernovae are expected. In parallel, we are developing LSTM-based models (Huber et al. 2022, 2024) to measure time delays between multiple images of different types of lensed transients from their multiband light curves, enabling precise and unbiased measurements with uncertainties below one day.

Agentic AI has the potential to transform scientific research by enabling AI systems to support complex, multi-step research workflows. We are among the early adopters of agentic AI in astronomy, using it under researcher supervision for tasks ranging from hypothesis generation and literature exploration to experiment design, execution, and analysis of results. Beyond research, we also explore AI-assisted teaching and education, including interactive demonstrations and other tools for communicating complex astronomical concepts.