Periodic Reporting for period 1 - GLADIUS (Gravitational Lensing Analysis for Data Intensive Upcoming Surveys)
Berichtszeitraum: 2020-10-01 bis 2022-09-30
Gravitational lensing is one of the major tools to address long-standing problems in astrophysics and cosmology, such as dark energy, dark matter, and black hole physics. Specifically:
I) Analysis of lensed quasars through time-delay cosmography provides an independent probe for the value of the Hubble constant, a crucial cosmological parameter for which different methods provide inconsistent results creating a tension in modern cosmology.
II) Dark matter properties within and around galaxies are still poorly understood. Gravitational lensing and microlensing constitute a unique way to measure them and test different dark matter models.
III) Quasars are crucial for understanding galaxy evolution and supermassive black hole growth. However, very little is known about the heart of the quasar: a supermassive black hole surrounded by a disc of accreting material. Rare microlensing events provide a unique resolving power, orders of magnitude higher than any current and foreseeable telescope, that can be used to measure the geometry and kinematics of quasar central regions.
Progress has so far been limited by the small number of known lenses and the intricate nature of the phenomenon, requiring painstaking handcrafted modelling approaches, hard-to-procure high-quality data, and lengthy computations. Upcoming all-sky surveys will revolutionize the field by discovering orders-of-magnitude more lenses and providing adequate information to understand them in depth. Mining these massive troves of data for the new astrophysical discoveries they hold requires a paradigm shift in our analysis approaches. GLADIUS undertook this task and produced innovative and flexible new methods merging traditional approaches and machine learning for the joint analysis of imaging and time-domain data. GLADIUS addressed the following three ambitious objectives:
1. Quantifying and mitigating the effect of microlensing on measuring the Hubble constant through time delay cosmography.
2. Creating a self-consistent lens modelling approach that fuses strong and microlensing models through a combination of traditional and machine learning methods.
3. Creating and deploying machine learning predictors using monitoring data from LSST, in order to provide alerts and optimal observation strategies for imminent microlensing high magnification events.
GLADIUS has achieved most of its objectives and milestones for the period, with relatively minor deviations. I have completed developing the method to quantify the effect of microlensing on measuring the Hubble constant, created two lens modelling methods that go beyond the state-of-the-art, and produced microlensing event predictors that are in the stage of fine-tuning before deployment within the LSST data processing pipeline.
I) Analysis of lensed quasars through time-delay cosmography provides an independent probe for the value of the Hubble constant, a crucial cosmological parameter for which different methods provide inconsistent results creating a tension in modern cosmology.
II) Dark matter properties within and around galaxies are still poorly understood. Gravitational lensing and microlensing constitute a unique way to measure them and test different dark matter models.
III) Quasars are crucial for understanding galaxy evolution and supermassive black hole growth. However, very little is known about the heart of the quasar: a supermassive black hole surrounded by a disc of accreting material. Rare microlensing events provide a unique resolving power, orders of magnitude higher than any current and foreseeable telescope, that can be used to measure the geometry and kinematics of quasar central regions.
Progress has so far been limited by the small number of known lenses and the intricate nature of the phenomenon, requiring painstaking handcrafted modelling approaches, hard-to-procure high-quality data, and lengthy computations. Upcoming all-sky surveys will revolutionize the field by discovering orders-of-magnitude more lenses and providing adequate information to understand them in depth. Mining these massive troves of data for the new astrophysical discoveries they hold requires a paradigm shift in our analysis approaches. GLADIUS undertook this task and produced innovative and flexible new methods merging traditional approaches and machine learning for the joint analysis of imaging and time-domain data. GLADIUS addressed the following three ambitious objectives:
1. Quantifying and mitigating the effect of microlensing on measuring the Hubble constant through time delay cosmography.
2. Creating a self-consistent lens modelling approach that fuses strong and microlensing models through a combination of traditional and machine learning methods.
3. Creating and deploying machine learning predictors using monitoring data from LSST, in order to provide alerts and optimal observation strategies for imminent microlensing high magnification events.
GLADIUS has achieved most of its objectives and milestones for the period, with relatively minor deviations. I have completed developing the method to quantify the effect of microlensing on measuring the Hubble constant, created two lens modelling methods that go beyond the state-of-the-art, and produced microlensing event predictors that are in the stage of fine-tuning before deployment within the LSST data processing pipeline.
The work performed during GLADIUS resulted in a combined treatment and understanding of strong lensing and microlensing. In order to achieve GLADIUS’ objectives, I developed four novel methods:
MOLET is a forward simulations code that generates mock imaging data for any kind of galaxy-scale lens potential, source, and instrument. A variable source, like a quasar or a supernova, can be included in the model, allowing for the simultaneous generation of light curves taking into account self-consistently time delays, microlensing variability, and other effects. Combining MOLET with PyCS I have quantified the effect of microlensing on measuring time delays.
VKL and Herculens are two lens modelling codes that go beyond reconstructing smooth, analytic forms of the lens mass distribution by including small scale perturbations. These can originate from different phenomena involving baryons and their interactions with dark matter, but recovering them heavily depends on prior assumptions. The codes I developed allow to investigate this by including novel, observationally justified priors, data-driven priors based on wavelet transforms and sparsity, and implicit neural network regularization. All methods have been tested on complex mock data (produced by MOLET).
Minotaur is a machine learning algorithm that can predict an imminent microlensing event in some window of time in the future based on LSST data extending some time in the past. It has been developed and tested on mock light curves produced by MOLET. I am currently expanding the code to cover a wide variety of possible cases and implementing it in the LSST pipeline.
In addition to these codes, I have created a neural network inspired by sound generation algorithms, like Google’s WaveNet, that can generate microlensing light curves conditioned on the target mass parameters of the lens and a custom GPU kernel to compare them to real data, which is x10 faster than a CPU. Once complete, this code will be merged with the VKL and Herculens codes to provide a joint macro- micro- lensing modelling method.
During the period of the project I was actively involved in the strong lensing science teams within the Euclid and LSST collaborations, which allowed me to take the first steps in implementing GLADIUS product methods to the Euclid and LSST pipelines and disseminate its results on image and light curve modelling.
GLADIUS resulted in 9 papers published in high-impact astronomical journals and 3 that are under review. I was invited to contribute to 4 review papers as part of the ‘Strong Gravitational Lensing’ workshop at ISSI Bern, Switzerland, and invited as a panelist to IAU’s symposium on machine learning, Busan, South Korea where I also presented my work. I organized the two ‘Lensing Odyssey’ 5-day workshops in Kouremenos, Greece (2021 and 2022), with 30 participants that resulted in new collaborations, proposals, and papers being written.
MOLET is a forward simulations code that generates mock imaging data for any kind of galaxy-scale lens potential, source, and instrument. A variable source, like a quasar or a supernova, can be included in the model, allowing for the simultaneous generation of light curves taking into account self-consistently time delays, microlensing variability, and other effects. Combining MOLET with PyCS I have quantified the effect of microlensing on measuring time delays.
VKL and Herculens are two lens modelling codes that go beyond reconstructing smooth, analytic forms of the lens mass distribution by including small scale perturbations. These can originate from different phenomena involving baryons and their interactions with dark matter, but recovering them heavily depends on prior assumptions. The codes I developed allow to investigate this by including novel, observationally justified priors, data-driven priors based on wavelet transforms and sparsity, and implicit neural network regularization. All methods have been tested on complex mock data (produced by MOLET).
Minotaur is a machine learning algorithm that can predict an imminent microlensing event in some window of time in the future based on LSST data extending some time in the past. It has been developed and tested on mock light curves produced by MOLET. I am currently expanding the code to cover a wide variety of possible cases and implementing it in the LSST pipeline.
In addition to these codes, I have created a neural network inspired by sound generation algorithms, like Google’s WaveNet, that can generate microlensing light curves conditioned on the target mass parameters of the lens and a custom GPU kernel to compare them to real data, which is x10 faster than a CPU. Once complete, this code will be merged with the VKL and Herculens codes to provide a joint macro- micro- lensing modelling method.
During the period of the project I was actively involved in the strong lensing science teams within the Euclid and LSST collaborations, which allowed me to take the first steps in implementing GLADIUS product methods to the Euclid and LSST pipelines and disseminate its results on image and light curve modelling.
GLADIUS resulted in 9 papers published in high-impact astronomical journals and 3 that are under review. I was invited to contribute to 4 review papers as part of the ‘Strong Gravitational Lensing’ workshop at ISSI Bern, Switzerland, and invited as a panelist to IAU’s symposium on machine learning, Busan, South Korea where I also presented my work. I organized the two ‘Lensing Odyssey’ 5-day workshops in Kouremenos, Greece (2021 and 2022), with 30 participants that resulted in new collaborations, proposals, and papers being written.
The successful completion of GLADIUS has allowed to establish myself as a world-leading expert in gravitational lensing analysis, focusing on its impact on fundamental questions about dark matter, dark energy, and black hole physics. I am poised to apply GLADIUS’ findings and tools to the upcoming big data era in astronomy, led by surveys like Euclid and LSST, and exploit this opportunity to address major open questions in astrophysics:
1. MOLET simulations enable mitigating the effect of microlensing on time-delay measurements and producing more precise results from time delay cosmography.
2. VKL and Herculens go well beyond the state-of-the-art, addressing the problem of priors in lens mass reconstructions in a multitude of new and better ways.
3. Minotaur is the first quantitative microlensing event predictor that will be deployed within the LSST data processing framework. Its predictions will be exploited by higher cadence and multi-wavelength follow-up observations.
Investing in big data analysis frameworks and machine learning techniques, which transcend astronomy and extend to other disciplines and industry, has increased my potential to carry out interdisciplinary research. Through GLADIUS I have expanded my professional network by consolidating ties with the groundbreaking Euclid, LSST, and TDCOSMO collaborations in Europe and USA. The acquired professional maturity and independence, as well as mentoring and managing skills, will eventually enable me to run my own research group.
1. MOLET simulations enable mitigating the effect of microlensing on time-delay measurements and producing more precise results from time delay cosmography.
2. VKL and Herculens go well beyond the state-of-the-art, addressing the problem of priors in lens mass reconstructions in a multitude of new and better ways.
3. Minotaur is the first quantitative microlensing event predictor that will be deployed within the LSST data processing framework. Its predictions will be exploited by higher cadence and multi-wavelength follow-up observations.
Investing in big data analysis frameworks and machine learning techniques, which transcend astronomy and extend to other disciplines and industry, has increased my potential to carry out interdisciplinary research. Through GLADIUS I have expanded my professional network by consolidating ties with the groundbreaking Euclid, LSST, and TDCOSMO collaborations in Europe and USA. The acquired professional maturity and independence, as well as mentoring and managing skills, will eventually enable me to run my own research group.