My research focuses on data-driven astronomy, combining large-scale survey observations with advanced machine learning methods. By analyzing supernovae, and galaxy clusters, I aim to uncover the fundamental nature and evolution of the Universe.
My research focuses on data-driven astronomy, combining large-scale survey observations with advanced machine learning methods. By analyzing supernovae, and galaxy clusters, I aim to uncover the fundamental nature and evolution of the Universe.
Key Projects
1. Cosmological Inference from Supernovae Using SBI and FPCA
I am developing a simulation-based inference framework to constrain cosmological parameters using multi-band photometry of Type Ia supernovae. Functional principal component analysis (FPCA) is used to characterize light curves, capturing a greater degree of temporal variations. The method accounts for host galaxy effects, including star formation rate, mass, and morphology, to obtain more precise estimates of key cosmological parameters.
2. FPCA-Enhanced Ensemble Learning for SNe Classification
I developed a fully data-driven method to distinguish Type Ia supernovae from other astrophysical transients using Functional Principal Component Analysis (FPCA). FPCA decomposes light curves into principal components that capture complex variability beyond template-based models like SALT2. I repurposed these component scores as features for classification, using ensemble-based machine learning. Trained on simulated LSST data and validated on the DES Y5 spectroscopic sample, the model achieves ~90% accuracy and ~95% precision, demonstrating robust cross-survey performance.
3. Cosmological Inference from Galaxy Clusters Using SBI
I have worked on galaxy clusters as an independent cosmological probe, constraining parameters using simulation-based inference (SBI). I built simulated cluster catalogs with analytical halo mass functions and richness–mass models to develop and validate the SBI framework, benchmarking it against state-of-the-art MCMC methods. SBI accurately recovers key cosmological parameters (Ωₘ, σ₈) within 2-σ limits, with constraints comparable to those from MCMC.
SBI offers key advantages over MCMC --- it does not require a closed-form likelihood function, marginalizes nuisance parameters during training, and supports fast, amortized inference on new datasets. (Paper)
4. Predicting Dark Matter from Baryonic Signatures Using ML
Dark matter subhalos play an important role in galaxy formation and evolution, but accurate prediction of dark matter properties remains a challenge of modern-day astronomy. Using Eagle simulations, I used machine learning framework (gradient boosted decision trees (GBDT) to predict the total mass and half-mass radius of dark matter subhalos from baryonic tracers (gas, stars, black holes). Gas radius and black hole mass emerge as the most important features for subhalo radius and mass estimation, respectively. The model achieves satisfactory performance and shows strong extrapolation to masses, redshifts, and simulation resolutions beyond the training range. This study highlights the potential of machine learning in advancing our understanding of dark matter. (Paper)
5. Classifying Galaxy Morphologies with Automated ML
Galaxies are the fundamental building blocks of the universe, consisting of stars, gas, dust, and dark matter halos. Their morphological classification provides key insights into how galaxies form and evolve over time. I used a number of machine learnring models to classify Sloan Digital Sky Survey (SDSS) samples into four categories—spirals, ellipticals, mergers, and stars.
Classifying mergers is particularly challenging because they are underrepresented in the dataset and share features with both spirals and ellipticals, creating a highly imbalanced learning problem. Starting from a large number of raw structural and photometric features, I applied Principal Component Analysis (PCA) to reduce dimensionality and used the 25 most significant components as model inputs. The models achieved a test accuracy of ~98%. To mitigate the class imbalance problem, we used oversampling to boost recall for underrepresented classes and conducted feature sensitivity tests to identify the most dominant parameters.
This study demonstrates that machine learning can capture complex morphological signatures, making it a powerful tool for next-generation galaxy surveys. (Paper)
6. Ensemble Tree-Based ML for Galaxy Photo‑z
I developed a machine learning method using ExtraTreesRegressor to estimate photometric redshifts of galaxies and quasars across the full SDSS range (0<z<7). Most previous studies have focused on z<1, and a few have used deep learning to estimate redshifts up to z<4, limited by the sparsity of high-redshift samples. To partially mitigate class imbalance, I transformed redshift values into the logarithmic domain. Furthermore, I introduced a custom scoring metric, redshift-weighted mean squared error, to penalize errors at higher redshifts more heavily and applied post-processing methods to compensate for model bias at high redshift values. I used 30 photometric and morphological features, applying recursive feature elimination to retain the most informative ones, and evaluated performance on a uniform test set divided into seven equally spaced bins across the full redshift range. The model achieved a mean squared error of 0.66, demonstrating robust performance across the entire redshift range. (Paper)