Research 

My research primarily focuses on cosmology, with a strong emphasis on using machine learning as a tool to detect new physics. Specifically, I have explored several areas, including applying machine learning to study non-trivial cosmic topology and the anomalies of the cosmic microwave background. Additionally, with collaborators, I have worked on optimizing laboratory tests of screened modified gravity by using numerical methods and genetic algorithms. My work also encompasses cosmological and astrophysical tests of gravity. Finally, I have also done work outside the confines of astrophysics and cosmology. Namely, recently I have co-authored a study on applying artificial intelligence as a tool to detect illegal ivory product sales on online platforms. 

To find out more about the mentioned research areas, click on the links below.

Cosmic Topology and the Anomalies of the Cosmic Microwave Background

Cosmic topology is the study of the overall shape and topological properties of the Universe. While cosmology typically deals with the local geometrical properties of the Universe, cosmic topology addresses questions about global properties, such as whether the Universe is finite or infinite, simply connected or multiply connected, closed or open. Non-trivial topology is expected to lead to potentially observable signatures in cosmological and astrophysical data, such as multiple images of cosmic objects (clone images) and matched-circle pairs in the cosmic microwave background (CMB). Previous observational searches, such as the search for matched-circle pairs in the CMB or cosmic crystallography, have allowed researchers to place competitive constraints on the size of the Universe.  

Figure 1: Matched-circle pairs in the CMB. The brown circles show the locations of the pairs of matched CMB fluctuation patterns appearing due to a universe having a 3-torus topology (Planck collaboration, 2013).

The studies of the CMB in the last several decades have exposed a number of unusual features in the data, that are arguably not expected to be observed in the standard model of cosmology. Starting with COBE and then later in WMAP and Planck mission data, numerous (generally large scale) anomalies have been observed in the CMB. This includes, but is not limited to the lack of power on the largest angular scales (low quadrupole and octupole moments), unexpected alignments of the largest-scale temperature fluctuations in the CMB (often referred to as Axis of Evil), or the deviations from the predicted statistical uniformity of the CMB (e.g, the so-called cold spot anomaly). In theory, many models with non-trivial topology could  offer possible explanations to many of the outlined anomalies. Namely, certain topologies can naturally suppress temperature fluctuations on large angular scales. Similarly, a multi-connected universe could impose geometric constraints that cause large-scale CMB fluctuations to align in unexpected ways. These and similar questions related to cosmic topology are currently being explored by the recently established COMPACT (Collaboration for Observations, Models and Predictions of Anomalies and Cosmic Topology) collaboration, which I am a member of. 

Figure 2: CMB temperature auto-correlation between different multipole values in the 3-torus (E1) topology of two different sizes. The characteristic checkerboard pattern appears due to non-trivial topology. 

Detecting evidence of cosmic topology in the CMB temperature and polarization data is a key challenge for the COMPACT collaboration. Recent work suggests that non-trivial cosmic topology leads to non-diagonal correlations between the different multipoles in the CMB. As illustrated in Fig. 2, this corresponds to the characteristic checkerboard pattern in the temperature auto-correlation matrix. Different topology classes present different correlation patterns. A key challenge, however, is to detect such correlations for topologies larger than the diameter of the last scattering surface. A universe with a non-trivial topology of such or larger size would present such correlations but would not have the characteristic matched-circle pair signature (Fig. 1), making it more difficult to detect. In general, we found that the larger the topology scale is, the more difficult it is to detect in the CMB temperature and polarization data. This is where techniques such as machine learning can play a key role.

Most recently, I have been working on techniques that employ artificial intelligence to look for evidence of non-trivial cosmic topology in the simulated CMB temperature and polarisation data. More specifically, we have developed a number of machine learning algorithms to distinguish and classify harmonic space realizations of the CMB with trivial and non-trivial topology. This work is the first step towards developing a set of machine learning-based methods to detect evidence (if such evidence exists) of non-trivial topology in current and future data of the CMB. Furthermore, techniques like this will give us further insights into the anomalous features of the CMB.

Relevant publications: 


The work of the COMPACT collaboration in the science news:

Optimizing Laboratory Tests of Modified Gravity

Models of modified gravity propose changes or extensions to the theory of general relativity (GR) to explain phenomena such as dark matter and dark energy. Modified gravity model building is generally difficult—a successful model has to satisfy numerous theoretical and observational constraints while being consistent with GR. One of the key observational challenges is meeting the stringent solar system constraints (e.g., from lunar laser ranging or Shapiro delay measurements) while still exhibiting interesting new physics effects. In this regard, an important class of models are scalar-tensor theories that possess a screening mechanism, which refers to the ability to turn off the modified gravity effects in high-density environments. Important examples of such theories include the Chameleon and Symmetron models.

Testing models with screening mechanisms is generally difficult, precisely because such theories are designed to subdue the modified gravity effects in high density regions. To overcome this, one can look for modified gravity effects (e.g., for a fifth force or, more generally, for deviations from GR) in vacuum chamber experiments. While there are many possible vacuum chamber experiment designs, they generally consist of a chamber with a source of a scalar field (eg., a small test mass in the center of the chamber) or a sensor optically held in place by a laser setup. A key question when it comes to designing such experiments is what would be the optimal shape of the aforementioned test mass or the vacuum chamber itself. While many vacuum chamber tests have considered spherical test masses, recent research suggests that non-spherical sources can generate a significantly stronger fifth force signal.  

Figure 1: A candidate for the optimal shape of a test mass for a vacuum chamber experiment. The right half of an axis-symmetric chamber is shown. The figure on the right shows the zoomed-in region of the test mass. The color refers to the value of the Chameleon fifth force (Briddon et al. 2024). 

Recently, we have addressed this question using techniques from machine learning. In our recent publication, we have developed a code that uses genetic algorithms in order to determine the shape of the test mass that maximises the fifth force for the model of Chameleon gravity. The code is based on a numerical finite-element-based solver, SELCIE, that allows finding the Chameleon fifth force profile for density sources of arbitrary shapes.  In combination with genetic algorithms, our numerical code allows us to determine the optimal shape of a test mass (for a set mass, symmetry settings, and a fixed fifth force measurement distance from the source). Fig. 1 above shows a candidate for an optimal shape of the test mass for a spherically symmetric vacuum chamber, with the red "x" marking the position at which the fifth force is maximized. 

Relevant publications: 

Astrophysical and Cosmological Tests of Gravity

Theories of modified gravity that possess screening mechanisms offer a possible path towards solving some of the key problems in modern cosmology. Specifically, modified gravity could provide crucial insights into understanding dark energy and dark matter. Exploring the theoretically and observationally allowed modifications of GR could give us clues on possible paths to more complete theories of gravity, such as quantum gravity. Given the plethora of available models, a natural question is whether cosmological and astrophysical data can be used to constrain and ultimately rule out different classes of theories.

To answer this question, I have recently been working on different methods of using cosmological and astrophysical measurements to constrain theories of modified gravity. In collaboration with astrophysicists studying galaxy clusters, we have developed a technique that combines X-ray surface brightness and weak lensing measurements by stacking data from hundreds of galaxy clusters. These measurements gauge the underlying density distribution in galaxy clusters, allowing us to compare the predictions from GR to those from different theories of gravity, such as f(R) and Chameleon gravity. An adaptation of this technique has also been used to constrain the model of emergent gravity.

Figure 1: Chameleon acceleration in relation to the virial mass and the concentration parameters for The Three Hundred Project galaxy clusters. The absolute value of the acceleration is shown on the left, while the ratio with the corresponding Newtonian acceleration for each cluster is show on the right (Tamosiunas et al. 2021).

A related question concerns the shapes of galaxies and galaxy clusters and their role in testing theories of gravity. Specifically, which shapes of galaxies and clusters maximize the chances of detecting the fifth force in different modified gravity theories? We have recently addressed this question by analyzing hydrodynamic simulations of galaxy clusters from The Three Hundred Project dataset. Specifically, we examined how the size of the fifth force in Chameleon gravity depends on the density distribution and the ellipticity of these galaxy clusters. Fig. 1 above shows the size of the Chameleon acceleration for the Three Hundred Project galaxy clusters in relation to their concentration and virial mass parameters. While we generally find the Chameleon acceleration to be small, we also observe a clear relationship between virial mass, concentration, and the corresponding acceleration. A similar relationship is observed for the cluster shape (ellipticity) and the Chameleon acceleration, which you can learn more about here.

Figure 2: The acceleration ratios for f(R) gravity, where we vary the different parameters that characterise the density distribution in cosmic voids. The corresponding ratios are also equal to the ratio of the fifth force and the corresponding Newtonian force in the mentioned model (Tamosiunas et al. 2022). 

The results above indicate that modified gravity effects are generally suppressed in high-density regions, such as galaxies and galaxy clusters. A natural question then is, what would be the optimal astrophysical system to maximize the fifth force? One possible answer is cosmic voids. Cosmic voids are the most under-dense regions in the cosmic web, offering a natural setting to look for the effects of modified gravity. Being the emptiest regions in the large-scale structure of the Universe, cosmic voids serve as a cosmic equivalent of the vacuum chambers described above. Recently, we have applied our numerical codes to investigate Chameleon screening in the context of cosmic voids. Specifically, we examined realistic cosmic void density profiles determined from CMB data, the SDSS survey, and N-body simulations. For these profiles, we numerically calculated the corresponding acceleration in Chameleon and f(R) gravity models and compared it against the Newtonian acceleration for the same profiles. Finally, we varied different void profile parameters to investigate how the fifth force (or acceleration) depends on the features of these profiles. Fig. 2 summarizes the results for f(R) gravity, where we vary two key parameters that determine the depth and density distribution in cosmic voids.

In summary, while a large part of the Chameleon parameter space is already ruled out by observational constraints, cosmic voids are evidently one of the best environments to test screening mechanisms, with voids of different depths giving rise to potentially measurable fifth force. Further research is required to determine the optimal shape of a cosmic voids for such fifth force searches.

Relevant publications:

Generative Models as Tools to Emulate Cosmological Data

Generative models are a class of algorithms in machine learning that generate new and unique data samples from the learned distribution from a given dataset. While various types of generative models have existed for some time, recent advances in machine learning have led to the development of new and more powerful algorithms, such as generative adversarial networks (GANs) and variational autoencoders (VAEs). These algorithms refer to different types of neural networks trained in a specific manner, allowing for the generation of statistically realistic and novel data, generally quickly and efficiently. For instance, after GANs were introduced in Goodfellow et al. 2014, the initial results of artificially created images were soon followed by breathtaking high-definition photographs of (non-existent) human faces that were essentially indistinguishable from real photos. Many other applications, including the generation of scientific data, have been developed since then. 

In our work, we are specifically interested in applying GANs as a tool to quickly and efficiently emulate cosmological simulations and weak lensing data. One of the key challenges in simulating the large-scale structure is the amount of time and computational resources required to generate a large number of datasets. High-resolution N-body and hydrodynamical simulations often require days or even weeks to run on high-performance computers. Additionally, creating high-resolution weak lensing maps requires extra steps, such as ray-tracing, to determine how light is affected by the simulated large-scale structure. This is where generative algorithms, such as GANs, offer a computationally less expensive alternative. A GAN can be trained on a relatively small dataset of cosmological simulation data, which can then be used to emulate realistic and novel data quickly and efficiently.

Figure 1: A comparison of randomly selected cosmic web slices from the training and the GAN-generated datasets. Columns 1 and 3 correspond to redshift z = 0.0 and columns 2 and 4 to redshift z=1.0 (Tamosiunas et al. 2021). 

Fig. 1 above shows several randomly selected samples from the training dataset we used to train a GAN algorithm to generate realistic 2D slices of the cosmic web (dark matter density field). We can also see the corresponding samples generated by our GAN code after the training procedure. While the generated data looks impressive to the eye, we conducted a thorough investigation of various statistics, including the power spectrum and Minkowski functionals, which show that the training and generated datasets are nearly indistinguishable, exhibiting only minor resolution-dependent differences in the Euler characteristic Minkowski functional. A fascinating feature of generative algorithms like GANs is their ability to generate novel data samples with cosmological parameters not present in the training dataset, through a process known as latent space interpolation. For instance, we managed to generate realistic cosmic web samples corresponding to a redshift z = 0.5 by interpolating between samples with redshifts z = 0.0 and z = 1.0 (similar to interpolating between different celebrity faces, as illustrated here). Features like these make GANs particularly interesting as tools for both generating novel scientific data and understanding how neural networks encode information during training. Our plan is to explore these questions further in our future work. 

Relevant publications:

Machine Learning Applied to Wildlife Conservation

Throughout my graduate studies and later as a postdoctoral researcher, I had the privilege of applying my scientific knowledge beyond the usual constraints of theoretical physics. I had the opportunity to work on several projects where I could apply my expertise in physics and machine learning to a wide array of problems. For example, my colleagues at the University of Portsmouth and I explored ways to use machine learning as a tool in wildlife conservation. In collaboration with experts from the faculties of business and law, we developed a machine learning algorithm to detect illegal ivory sales on online platforms. Despite most ivory products being illegal, they are still being sold and purchased on platforms such as eBay. What makes it particularly challenging to stop such illegal trading is that these ivory products are often disguised as similar but legal items, such as ivorine (a synthetic material resembling ivory) or ox bone. Furthermore, illegal sellers frequently use coded language and keywords to disguise the true nature of the materials being sold. This underscores the need for efficient and reliable automated algorithms that can detect these keywords in advertisements and analyze images of the products being sold.

Figure 1: A selection of carved ivory images along with their pixel-level classification obtained using image segmentation. The colors indicate whether a given pixel corresponds to ivory (lighter), bone (darker) or background (white) (Mozzon et al. 2024).

As a first step in tackling this problem, we developed an image segmentation approach that allows for pixel-level image classification. Based on the original U-Net neural network architecture, our algorithm can classify each pixel of an input image into three classes: ivory, bone, or background. We find that the algorithm shows promise in correctly distinguishing ivory from bone, based on a training dataset containing only several hundred public domain images. Nonetheless, our study also highlights some key challenges for such an image segmentation approach. Specifically, we find that the results depend strongly on the lighting conditions in each image (e.g., bright areas or reflections can lead to incorrect classification). We hope to address these issues in further studies with significantly larger datasets.

The pilot study outlined above is just a first step toward developing a set of algorithms that we hope will become an efficient tool for law enforcement in combating illegal wildlife trade. A natural next step is to also integrate the textual information available on platforms such as eBay. With the recent advancements in large language models, we can expect to significantly improve the results obtained when using image data alone. If proven effective, algorithms like these are expected to become part of the infrastructure of online trading platforms, used for flagging products sold illegally.

Relevant publications: