I'm an M.Sc. Astronomy student at IIT Indore, working toward computational astrophysics. I write
numerical and machine-learning code for astrophysical problems, and I test whether a model is learning
the physics or a shortcut.
My interest is where data-driven models meet physical understanding. Machine learning is
increasingly used to forecast astrophysical and space-weather events, but a high score only means something
if the model is predicting the event rather than recognising its first moments. Much of my work is about
designing tests that tell those two apart.
Alongside that, I'm building my foundation in numerical astrophysics through my M.Sc. courses
in Astrophysical Fluids and Plasma and Numerical Techniques for Astronomy. I want to work where
simulation, observation and learning meet.
I introduce a model-agnostic guard-gap cross-validation test that measures how much of a
flare-forecasting classifier's score comes from reading the flare onset inside its input window, rather
than predicting it. Applied to an LSTM forecaster on 80 M dwarfs (TESS 20-s cadence, leave-one-star-out),
a 10-minute guard cut the robust mean ROC-AUC from 0.997 to 0.870, against a within-star permutation
floor of 0.492 ± 0.026. The final ten minutes carry about a quarter of the above-chance signal.
Figure 1. What onset leakage looks like
A simulated M-dwarf light curve, about four hours across, with flares in red.
Point at it to place a forecaster's 120-minute input window. If the last ten minutes of the window touch a
flare's rise, the model can see the flare instead of predicting it.
Figure 2. How much of the score is forecasting?
Robust mean ROC-AUC, leave-one-star-out; the dashed line is chance. On the un-guarded task, a logistic
regression on 16 summary statistics, with no temporal modelling, reaches 0.998 and matches the LSTM:
evidence that the benchmark was measuring onset detection, not forecasting.
Share of the above-chance signal carried by the last ten minutes:
(0.997 − 0.870) / (0.997 − 0.492) = 0.127 / 0.505 ≈ 0.25
The final ten minutes carry about a quarter of the above-chance signal. The rest survives the guard.
I built an XGBoost baseline and an LSTM model with an intensity-weighted MSE loss to predict the
storm-time EEJ over the Indian sector from NASA OMNI solar-wind data. The full pipeline is deployed as a
FastAPI REST service with a web dashboard.
SYM-H predicted 60 minutes ahead
—nT
Fetching the latest prediction…
~32%
MAE improvement on intense storms
−518 nT
SYM-H in the May 2024 Solar Cycle 25 extreme event, a documented domain-shift failure
The May 2024 event lies far outside the range the model saw in training. I document it as a failure case
rather than hide it: knowing where a model breaks is part of the result.
The Sod shock tube is the standard first test of a hydrodynamics code: high pressure on the
left, low on the right, released at t = 0. Its exact solution, a rarefaction, a contact
discontinuity and a shock, is known, so every scheme can be checked against it.
Figure 3. Numerical solution against the exact Riemann solution
Exact solutionNumerical cellst = 0.000 of 0.200L1 density error —0 steps
What to try
Raise the cell count or switch to 2nd order and watch the L1 error fall.
What to notice
The contact smears the most, because no pressure jump holds it together.
Coming next
My PLUTO jet simulation, and an orbital integrator comparing leapfrog with RK4 energy drift.
Projects
Other things I've built.
Oct 2023
NASA International Space Apps Challenge
Exoplanetary environment simulation, as team leader
Led a team of four to design and simulate 3D exoplanetary environments using NASA's open data
Modelled conditions for potential extraterrestrial habitability using physics-based algorithms
2024
Educational outreach
Light-to-Mars model
Developed an Arduino-based interactive model using LEDs to visualise the speed of light and interplanetary distances
Designed to simplify complex astrophysical concepts for students
Experience
Where I've trained.
Jan 2025
CSIR–National Physical Laboratory, New Delhi
Trainee, Optical Radiation Metrology Training Programme
Completed a training programme on optical radiation metrology at CSIR-NPL
Gained hands-on experience in instrumentation and experimental data collection
2024
Starscapes Experience, Gurgaon
Intern, Product Development and Outreach
Assisted in astrophysical data analysis and visualisation using Python and MATLAB
Collaborated on developing astronomy outreach material for diverse audiences
2024
Nehru Planetarium, New Delhi
Participant, Astronomy Bootcamp
Attended workshops on cosmology, astrophysics and general relativity
Developed Python-based simulations for astronomical data interpretation
Education
Where I've studied.
2026 – present
M.Sc. Astronomy
Indian Institute of Technology Indore
2022 – 2025
B.Sc. Physics (Hons), First Division
Shivaji College, University of Delhi
2022
Senior Secondary, 90.0%
CBSE
2020
Secondary, 94.2%
CBSE
Key courses
Astrophysics and computation (M.Sc.)
Astrophysical Fluids and Plasma; Numerical Techniques for Astronomy
Physics (M.Sc.)
Classical Mechanics; Classical Electrodynamics; Quantum Mechanics and Spectroscopy
Physics (B.Sc.)
Mechanics; Waves and Oscillations; Electricity and Magnetism; Thermal Physics; Light and Matter; Modern Physics; Solid State Physics; Electromagnetic Theory; Quantum Mechanics I; Statistical Mechanics; Atomic, Molecular and Nuclear Physics; Research Methodology
Electronics
Electrical Circuit Analysis; Analog Electronics; Digital Electronics; Communication Systems
Programming with Python; Data Analysis and Visualisation using Python; Database Management Systems; Introduction to Web Programming
Achievements and certifications
2026
All India Rank 465, IIT JAM 2026, for Astronomy admission to IIT Indore
Oct 2025
World Space Week 2025 participation, Propagation Institute for Advanced Studies
Feb 2025
Arduino Training (score 90%), Spoken Tutorial Project, IIT Bombay
Jan 2025
Optical Radiation Metrology Training Programme, CSIR-National Physical Laboratory, New Delhi
Jun 2024
Astronomy: Exploring Time and Space, Coursera, University of Arizona
May 2023
IAAC 2023 (International Astronomy and Astrophysics Competition), qualified for the pre-final round
2018–19
District Topper, Vidyarthi Vigyan Manthan (VVM) 2018–19, national level
Skills
Programming
Python, C++, JavaScript, MATLAB, MySQL
Data and machine learning
Data analysis, XGBoost, LSTM
Web and deployment
HTML, CSS, FastAPI, Cloudflare Pages, GitHub
Tools and systems
LaTeX, Linux (Ubuntu), Nextcloud, Tailscale, Cloudflare Tunnels
Languages
English and Hindi (fluent), German (basic)
Research opportunities, questions, collaborations.
I'm looking for PhD positions in computational astrophysics. I'm always glad to talk about
numerical methods, simulations, machine learning for astrophysical data, or a project you think I'd enjoy.