Project title: Trustworthy machine learning for cosmological discovery
PI: Pedro Ferreira
As cosmological datasets grow increasingly detailed and offer new insights into large-scale cosmic structures and individual astrophysical processes, our theoretical models must adapt. This requires not only more advanced modelling of complex phenomena, such as galaxy formation, but faster and more scalable methods are needed to bridge the gap between predictions and observations. Machine learning (ML) presents a promising approach to accelerating these advancements, offering the potential to handle the immense computational demands of cosmological analysis. Indeed, the field of cosmology has become a rich playground for ML-enhanced and simulation-based analyses. However, the opaque or "black box" nature of ML models poses a significant challenge: how can we ensure that these models remain consistent with established physical principles? This lack of transparency can undermine the reliability of ML-enhanced analyses, making it difficult to trust the results. This proposal will develop trustworthy, high-performance ML models that adhere to physical frameworks, enabling the discovery of new Physics and enhancing our understanding of galaxy formation. This will bridge the gap between cutting-edge data and reliable, interpretable scientific conclusions.
