Scientific Projects

DeepCosmoNet

Artistic impression of Large Scale Structure of the Universe

Context

The DeepCosmoNet project aims to leverage Data Science, specifically Deep Learning, to advance our understanding of the large-scale structure of the universe, commonly referred to as the Cosmic Web.

Today, cosmology relies heavily on numerical simulations, especially N-body simulations, to study the large-scale structure of the universe. The DeepCosmoNet project has been selected and funded under the ICSC Programme, financed as part of the PNRR, funded by the European Union – NextGenerationEU.

Challenge

These simulations are not only computationally intensive but also require significant time for analysis. Specifically, identifying components of the Cosmic Web structure (such as halos, voids, and filaments), from 3D point-cloud data that result from cosmological simulations simulation is a time-consuming task. This lengthy process significantly slows down the pace of scientific research.

Various tools like SUBFIND, ZOBOV, Popcorn, and SCMS are currently used to identify different components of the Cosmic Web. However, each tool specialises in identifying only one type of component and the analysis process remains time-consuming.

Solution

DeepCosmoNet aims to develop specialised Machine Learning (ML) models for each Cosmic Web structure component that are significantly faster than current state-of-the-art tools. Additionally, the project will create an integrated pipeline capable of identifying multiple types of components simultaneously.

To achieve optimal performance, Neural Architecture Search (NAS) techniques will be investigated and implemented to optimise these ML models.

Main Outcomes

Advanced Segmentation

Development of machine learning models specifically designed for the automatic segmentation of complex 3D point-cloud data. These models will identify Cosmic Web structures such as halos, filaments, and voids within cosmological numerical simulations. Training will occur using datasets labelled with existing catalogues, with the primary goal of significantly reducing the time required for identifying cosmic structures, allowing researchers to explore more models in less time.

Cosmological Segmentation Pipeline

Creation of a comprehensive cosmological segmentation pipeline that leverages the advanced segmentation models developed previously. This pipeline will support the simultaneous detection of various cosmic structures, eliminating the need for separate tools dedicated to individual components. This approach will further decrease the time researchers spend on non-scientific activities.

Optimise Analysis

Use of Neural Architecture Search (NAS) techniques to enhance the efficiency of identifying the optimal neural networks for the machine learning models, including 3D Convolutional Neural Networks (3D CNNs) and Graph Neural Networks (GNNs). This optimisation will ensure that the models are effective and adaptable to varying requirements in cosmic structure analysis.

Koexai. Transforming Data into Smiles Koexai. Transforming Data into Smiles

Koexai’s data science team boasts an exceptional background, tackles challenges with enthusiasm, and provides valuable solutions.

Innovative and pragmatic data solutions driven by a passion for excellence and a thirst for knowledge

ContactsContacts