The Future of Knowledge

Modern telescopes take far more images than humans could ever look through, and so we rely on automated methods to measure what galaxies look like. AI is an excellent tool for solving this visual recognition problem. I’ll talk about how AI models have changed over the last few years to learn efficiently and quickly adapt to new measurement tasks.

I will discuss the uses of CNNs to identify gravitational lenses in simulated and real data, discussing common flaws that machine learning models suffer from and what can cause these issues. I will discuss the use of interpretability methods to identify what features within the image the machine learning techniques associate with gravitational lensing. I will explain how the need of interpretability in machine learning has led me to use U-nets to help locate gravitational lenses in images.

Gravitational lensing is a powerful tool that directly probes all clustering components in the universe through their gravitational effect on light from distant background sources. The problem arises in finding gravitational lenses, and, with the accelerated growth in data volume and complexity in astronomy, machine-learning-aided lens searches have proven successful.

We present a proof of concept for an alternative method of strong gravitational lens finding using a conditional Generative Adversarial Network (cGAN). We use Early Release Observation (ERO) images of the Perseus Cluster from Euclid, covering 0.57 sq. degrees on the sky, and the network is based on the pix2pix architecture with an adapted U-Net generator. We train our model to predict Euclid’s NISP-H band flux (1.54-2.00µm) from a combination of the filters NISP-J, NISP-Y and VIS band (0.55-1.54µm) in 40,000 cut-outs from the Perseus Cluster which are 20×20 arcseconds in size. We test the cGAN on 5,000 cut-outs from the Perseus cluster, 10% of which contain a simulated strong gravitational lens painted into the cut-out based on a Singular Isothermal Ellipsoid model. Candidate gravitational lenses and cut-outs with a gravitational lens painted in were deliberated excluded from the model’s training data set such that gravitational lensing systems remain unknown to the network. We find that the cGAN can accurately predict the NISP-H band flux of the cut-outs from the Perseus cluster. However, the model fails to predict the NISP-H band flux of the cut-outs containing the simulated gravitational lenses, with a larger difference between the model’s prediction and ground truth for lenses with extended arcs and Einstein rings, suggesting that the cGAN can be used as an anomaly detector for an alternative method of lens finding.

The upcoming decade promises remarkable progress in our grasp of exoplanet formation, evolution, and an in-depth assessment of their climates and possible habitability. With the James Webb Space Telescope now in orbit, we’re at the threshold of high-precision atmospheric studies of these distant worlds. This milestone, combined with forthcoming endeavours like the ELTs and the Ariel space mission, positions us in an era rich in high-precision exoplanetary data.
 
This vast influx of precise data holds immense discovery prospects but also introduces challenges in handling both model intricacy and data volume. As machine and deep learning techniques gain traction across various scientific and industrial domains, they also find their footing in exoplanetary and solar system research. Such techniques offer a refreshing edge in modelling intricate, non-linear data over conventional methods. In my talk’s initial segment, I’ll explore the use of simulation-based inference, deep surrogate models, and Explainable AI in refining time-series data and leveraging inverse modelling to extract exoplanet attributes from observations.

In the second part, I’ll delve into our machine-learning-centric data challenges. It’s noteworthy that many challenges in exoplanet data analysis resonate with issues keenly explored in the machine learning realm. Yet, interdisciplinary collaboration has often been stymied by language barriers and domain unfamiliarity. Addressing this, as part of the ESA Ariel Space mission, we have organised four machine learning challenges, hosted at prestigious AI conferences like NeurIPS and ECML, attracting hundreds of participating teams annually. I will discuss how such challenges can catalyse rapid, sturdy AI advancements in this budding domain and pave the way for more resilient AI solutions in the future.

The work is to develop a machine learning classifier (NEEDLE) for identifying superluminous-supernovae and tidal disruption events at their early stage, using a few detections and the context information. It is currently applied on Lasair broker digesting ZTF alerts, and in the future, it will deal with LSST alerts.

Machine learning has been widely used in different astronomy researches recently. Some of its models, especially the image to property feature, make it the perfect tool for solving some particular problems in astronomy, such as estimating the cluster mass from observation images. In this showcase, I present how the different ML models are adopted to estimate the cluster mass in different dimensions — from a signal cluster mass M_500 (0d data point, see de Andres et al. 2022), to cluster mass within different radii (1D density profile, Ferragamo et al. 2023) and projected cluster mass map (2D image, de Andres et al. 2024). With this powerful tool, we are able to probe the mystery Universe, deeply, completely and swiftly!

I provide an overview of recent work constraining cosmological parameters from two of the largest galaxy imaging surveys to date, the Dark Energy Survey and the Kilo-Degree Survey. Both analyses use a novel approach to inference: machine-learning the likelihood using neural density estimators and large suites of large-scale structure simulations. I’ll include brief introductions to weak gravitational lensing, on which the analyses are based, and to the concepts of simulation-based/likelihood-free inference.