I'm Dr. Magdalena Siwek

Computational Scientist

Stochastic modeling, time-series inference, and ML surrogates for expensive simulations.

About Me

About Me

I'm Magdalena Siwek, a computational scientist.

I build stochastic models and ML systems for high-dimensional, noisy time-series data, and ship them as production Python. My work spans time-series inference, probabilistic modeling, and ML surrogates for expensive simulations, with HPC experience at ~3M CPU-hour scale. I earned my PhD from Harvard in 2024, and since then I've been a Fellow with the Simons Society of Fellows, first at Columbia and now at New York University, where I built and released calypso — a stochastic time-series emulator that runs ~10³× cheaper at inference than a neural baseline. I remain a full member of the NANOGrav collaboration, and co-develop the population-synthesis code holodeck.

I am originally from Germany, and completed my undergraduate (BSc) and graduate (MSci) studies in physics and astronomy at the University of Glasgow before starting my PhD at Harvard in 2018. I was also a Teaching Fellow at Harvard (2021–22, Bok Center Distinction in Teaching). While at Harvard I founded and ran the CfA's community garden. I am also an avid mushroom hunter.

  • Name:Magdalena Siwek
  • Focus:Stochastic processes, time-series inference, ML surrogates, HPC
  • Current Position:Fellow, Simons Society of Fellows (Columbia & NYU)
  • Languages:Python (NumPy, SciPy, PyTorch, Pandas), C, Cython
  • Education:
    2024 - PhD in Astrophysics, Harvard University
    2018 - MSci, University of Glasgow
    2017 - BSc, University of Glasgow

Research

Research

Publications

Publications

2026

calypso: a Parameter-Conditioned Stochastic Surrogate Model for Circumbinary Accretion Time-Series

Magdalena Siwek et al.

A fully stochastic, interpretable emulator for expensive high-dimensional time series. A PCA + multivariate-Gaussian sampler interpolates across a 2-D parameter space, returning calibrated realizations at ~10³× lower inference cost than a neural baseline — no hyperparameter tuning, no training instability. Released as calypso-emulator on PyPI with a live Streamlit demo and an automated release pipeline.

October 2024

Signatures of Circumbinary Disk Dynamics in Multi-Messenger Population Studies of Massive Black Hole Binaries

Magdalena Siwek, Luke Zoltan Kelley, Lars Hernquist

In this work, we combine the CBD models from my suite of hydrodynamic simulations and apply it to a MBHB population synthesis model based on galaxy merger rates from the cosmological simulation Illustris. We find that MBHB populations detected in PTAs, LSST and even LISA show a significant (1-2 orders of magnitude) increase in their eccentricity distributions when CBD models are applied. Our results suggest that detections of eccentric MBHBs are the rule rather than the exception in upcoming transient surveys, provided that CBDs regularly form in MBHB systems. This is due to the eccentricity pumping effects found in my simulations. We also find that eccentric harmonics are much more likely to be detected in LISA with the influence of CBD accretion.

June 2023

Orbital Evolution of Binaries in Circumbinary Disks

Magdalena Siwek, Rainer Weinberger, Lars Hernquist

Binaries on many scales encounter circumbinary disk (CBD) driven evolution at some point in their lives. But how does the presence of a CBD affect the orbital elements of the binary? In this work, we ran the largest to-date parameter study over binary mass ratio (q) and eccentricity (e), evaluating how the interaction with the CBD is affected by varying the parameters q and e. We discovered that mass ratio and eccentricity become correlated very quickly in the presence of a CBD, that is, the eccentricity evolves to an equilibrium value that is determined by the mass ratio of the system. I am excited to see whether we can find evidence of this dynamical effect in massive black hole or stellar binary populations.