PhD Defense by Indira Ocampo Justiniano: A machine learning approach for unraveling the nature of dark matter and dark energy
However, the vast quantity and quality of the data collected from present and upcoming cosmological surveys has become so demanding that sophisticated computational tools are required. As a result, machine learning (ML) methods are being implemented for data analysis and model constraints, with the hope of alleviating tensions in the parameters inferred from different cosmological probes.
A significant challenge remains in the complexity of these models: as architectures become more accurate, this usually comes with the trade-off of their interpretability. This can lead to a lack of transparency and trustworthiness in a field that demands validation based on physical laws rather than artifacts arising during training.
The research presented in this thesis focuses on different ML architectures based on neural networks (NNs) to search for signatures of physics beyond ΛCDM. In Chapter 4, we train a convolutional neural network on the Quijote N-body simulations to perform parameter inference of fσ8.
In Chapter 5, we discuss the results of testing a type of Coupled Dark Energy model against ΛCDM, by training an NN to discriminate the simulated fσ8 data. We extended this approach in Chapter 6, by testing another alternative to ΛCDM, the f(R) Hu-Sawicki model, while this time we applied LIME (Local Interpretable Model-agnostic Explanations) after the NN was trained. LIME serves as an interpretability framework that assigns importance scores to input data, identifying the features most relevant to the model’s decision-making. For both Chapter 5 and Chapter 6, data simulations were tailored to the DESI survey specifications.
In Chapter 7, we applied a similar pipeline trained on Cosmic Microwave Background simulated temperature and polarisation power spectra. For this project, we targeted the Planck mission, simulating data coming from ΛCDM, f(R) and a non-standard oscillating feature in the primordial power spectrum. To explain individual model predictions, we implement SHAP (SHapley Additive exPlanations).
In Chapter 8, we extend this analysis by performing the training and interpretability analysis directly on the CMB maps, integrated into my publicly available package SkyExplain¹.
Finally, in the Conclusions and future challenges, I draw the main conclusions and discuss future work.