●RESEARCH IN PROGRESS · ACTIVE RESEARCH NOTEBOOK

Deep Learning the Cosmic Web

Reconstructing Invisible Dark Matter Structures via U-Net Architectures. An ongoing undergraduate research project investigating deep learning approaches to understand the large-scale structure of the cosmic web.

RESEARCH DOMAINAstrophysics & Deep Learning

PROJECT TYPEUndergraduate Research Project

PRIMARY FOCUS3D U-Net Spatial Reconstruction from Astronomical Data

TIMELINE & STATUS2026 – Present (In Progress)

01 / RESEARCH QUESTION & SCIENTIFIC CONTEXTPROBLEM STATEMENT

In modern cosmological physics, dark matter makes up the dominant matter content of the universe, forming a vast, filamentary web that guides galaxy formation and cluster evolution. However, dark matter does not interact with the electromagnetic spectrum and cannot be imaged directly with telescopes.

Astronomical surveys only directly observe luminous baryonic matter—such as galaxies and gas. This research investigates whether deep learning encoder-decoder architectures, particularly 3D U-Net variants, can learn non-linear spatial mappings to reconstruct underlying continuous dark matter density fields from observable astronomical tracer data.

02 / RESEARCH METHODOLOGY & INVESTIGATION WORKFLOWACTIVE NOTEBOOK

The research follows a systematic investigation pipeline currently in progress:

1. RESEARCH QUESTION [ACTIVE]→2. LITERATURE REVIEW [ACTIVE]→3. METHODOLOGY DESIGN [IN PROGRESS]→4. EXPERIMENTATION [PLANNED]→5. EVALUATION [PLANNED]

STAGE 03: U-NET ARCHITECTURE — Encoder-Decoder Feature Learning[Architecture design in progress]

RESEARCH OBJECTIVE

Investigate U-Net convolutional architectures for spatial pattern learning and multi-scale feature mapping.

DOMAIN METHODOLOGY

U-Net encoder-decoder architectures with skip connections allow networks to learn hierarchical representations while retaining fine spatial context.

PYTORCH / ARCHITECTURAL APPROACH

Implementing PyTorch U-Net models to study reconstruction fidelity across multi-scale spatial features.

INVESTIGATION SCOPE

Ongoing undergraduate research investigating U-Net architectures for dark matter structure reconstruction from astronomical data.

RESEARCH STATUS: Ongoing undergraduate project at University of Moratuwa.RESEARCH IN PROGRESS

03 / KEY INVESTIGATION AREASDEEP LEARNING CHALLENGES

Spatial Feature Preservation via Skip Connections

Investigating how U-Net skip connections allow localized high-frequency spatial details from the encoder path to bypass downsampling bottlenecks and assist the decoder in reconstructing thin cosmic filaments.

Astronomical Density Scaling & Transformations

Cosmological density fields span wide dynamic ranges between sparse voids and dense gravitational halos. Exploring density transforms and normalization techniques to stabilize network training.

Loss Formulation for Physical Consistency

Studying loss objectives that balance localized voxel reconstruction with large-scale spatial structural correlation.

Statistical Validation Framework

Developing statistical analysis routines in Python to evaluate cross-correlation and scale consistency between reconstructed fields and ground-truth simulated structures.

05 / CURRENT STATUS & RESEARCH NOTEHONEST DISCLOSURE

This project is an active undergraduate research investigation at the University of Moratuwa. In keeping with scientific integrity, experimental results and final statistical findings will be documented and published upon rigorous validation.