I'm an astrophysicist with a background in computer science,
specializing in solving complex problems across science and industry.
I develop AI-driven solutions for diverse, unconventional
datasets, combining deep domain understanding
with cutting-edge machine learning to unlock new insights.
About Me
Hi! I'm Rohan Pattnaik, an astrophysicist and applied AI specialist with a unique journey bridging Computer Science and Astrophysics.
My journey started with an undergraduate degree in Computer Science, where my curiosity for challenging problems sparked an interest
in astrophysics—a field full of intriguing data and unique problems.
My Journey
My career path is rooted in curiosity and interdisciplinary exploration. I've had the privilege of working on fascinating problems,
from analyzing astronomical data—like radio, X-ray, and optical spectra—to leveraging AI in industry sectors such as finance
and energy. Each new domain has deepened my appreciation for data, reinforcing my belief that quality data is equally—if not more—crucial than sophisticated algorithms.
What I Do
I excel at rapidly immersing myself in new fields, understanding their core challenges, and determining precisely how AI can make a meaningful difference.
My unique experience with unconventional datasets, like galaxy spectra, has equipped me with a distinctive approach to problem-solving. Currently,
I'm actively involved in collaborations within astrophysics research, contributing insights that streamline and enhance traditional methodologies.
Let's Collaborate!
Today, I actively advise astronomy research projects worldwide, including collaborations spanning my past institutions.
I’m always open to new challenges—especially those that require innovative, data-driven approaches that traditional
consulting may not offer. Explore my Skills & Research page for deeper insights into my work and to see
how my skills and experience might align with your projects. If you're interested in collaborating or have intriguing challenges you'd
like to discuss, please reach out via the Contact page.
Skills Summary
Programming Languages
Python
C++
R
Julia
Java
C
Bash
SQL-lite
Tools
Pytorch
Pandas
Keras
Numpy
Scikit-Learn
Git
HTML5
CSS3
Latex
Research & Projects
SpecPT (Spectroscopy Pre-trained Transformer) Model for Galaxy Spectra
Spectroscopy Pre-trained Transformer (SpecPT) is a transformer-based model developed for analyzing spectroscopic data,
including spectrum reconstruction and redshift estimation. Trained on the DESI Early Data Release (EDR), SpecPT demonstrates
strong performance on both Bright Galaxy Survey (BGS) and Emission Line Galaxy (ELG) samples. It accurately reconstructs
spectra—capturing emission lines, absorption features, and continuum shapes—while effectively denoising the input.
For redshift prediction, SpecPT achieves high accuracy with with Normalized Median Absolute Deviation (NMAD)
values of 0.0006 and 0.0008, and low catastrophic outlier fractions of 0.20% and 0.80%, respectively.
It performs reliably across the full redshift range (0<z<1.6), highlighting its robustness. SpecPT's
learned latent representations also enable downstream applications, such as outlier
detection, ISM property estimation, and transfer learning across datasets.
[Paper Link]
Redshift Wrangler: Citizen Science Spectroscopy
Redshift Wrangler is a Zooniverse-based citizen science project that invites the public to help measure galaxy
redshifts—key to understanding how galaxies evolve across cosmic time. By identifying spectral features in galaxy
spectra, volunteers contribute to locating galaxies on the timeline of the universe. I developed a custom pipeline
that converts raw FITS data into clean, interpretable images by dynamically adjusting contrast and brightness,
enabling accurate public annotations. These labeled spectra serve as foundational training data for future machine
learning models, enhancing both accuracy and interpretability through supervised and contrastive learning approaches.
[Check Out Redshift Wrangler on Zooniverse]
Classifying Low-Mass X-ray Binaries (LMXBs) using Machine Learning
Low-mass X-ray binaries (LMXBs) are binary systems where one of the components is either a black hole or a neutron star
and the other is a less massive star. It is challenging to unambiguously determine whether an LMXB hosts a black hole or
a neutron star. In the last few decades, multiple observational works have tried, with different levels of success, to
address this problem. We developed a machine learning approach to classify LMXBs based on their compact object nature using
a Random Forest classifier trained on 5-25 keV energy spectra from the RXTE archive, achieving ~87% classification accuracy.
We further use the trained model for predicting the classes for LMXB systems with unknown or ambiguous classification. For
more details check out our paper linked below.
[Paper Link]
Seasonal Model Robustness in Financial Forecasting
At J.P. Morgan, I worked on evaluating deposit forecast models used for regulatory compliance and risk management.
A newly introduced X-13 model, while sophisticated, showed signs of overfitting on simpler data. I designed a custom
cross-validation framework to assess model performance, which confirmed these concerns. My analysis led to retaining
simpler, more stable models—improving forecast accuracy while reducing overfitting risk in a high-stakes financial environment.
ResNet-based Rock Particle Classification at SLB
At SLB, I enhanced rock particle classification by tackling bias introduced by heuristic labels that compromised model accuracy.
First, I developed a secure, web-based image labeling app enabling field experts to annotate samples directly—improving label
quality and enabling scalable dataset expansion. Using this improved dataset, I fine-tuned a ResNet-based classifier, achieving
over 85% accuracy with minimal training data. This project delivered the first proof of concept for automated rock analysis at SLB.
The image labeling tool developed during this project is available on my GitHub.
[GitHub Project Link]
Real-Time Transient Classification
Developed CNN-based classifiers to identify transient events (e.g., supernovae) from real-time photometric data during a research
visit to Swinburne University. This automation reduced human inspection volume by over 95%, significantly improving the response
rate to time-critical events in astronomy.
Detecting the 21cm Reionization Signal
Developed an artificial neural network model to distinguish faint 21cm Epoch of Reionization signals from foreground
contamination using simulations. Achieved 92% accuracy and demonstrated the potential of ML for cosmological signal extraction,
advancing efforts in observational cosmology.
[Paper Link]
