CV

Andrés Vicente Arévalo — PhD astrophysicist and ML researcher. Graph neural networks, inverse problems, HPC. Experience, results, education and publications.
Andrés Vicente Arévalo

Andrés Vicente Arévalo

PhD · Astrophysicist · ML researcher
[email protected] github.com/andreuva ORCID 0000-0003-3896-836X Freiburg, Germany · from Valencia, Spain · EU citizen
Core skills
Python / scientific stack10 yrs
Inverse problems & optimisation6 yrs
Deep learning (PyTorch, GNNs)6 yrs
HPC · MPI · GPU clusters7 yrs
C++ / Fortran numerics6 yrs
Data pipelines & reduction8 yrs
Tools
PythonPyTorchPyTorch GeometricHugging FaceW&BNumPy / SciPyC++CUDAFortranMPISLURMGitLinuxLaTeX
Compute
14.5MCPU-hours as PI/user
7.5M La Palma · 5M MareNostrum (BSC) · 2M HLRS
Languages
Spanish, CatalannativeEnglishfluentGermanbasic · learning
Download CV (PDF)

Astrophysicist with ten years of turning imperfect measurements into physical answers. I co-developed the first fully 3D non-LTE inversion of a solar prominence, and built graph-neural-network surrogates that run the expensive physics 10³–10⁶ times faster. Seeking research-scientist, data-science and machine-learning roles where expensive models meet messy data.

01Experience

2025 – nowPostdoctoral Researcher
Leibniz Institute for Solar Physics (KIS), Freiburg · DFG-funded
  • Postdoc in the DFG project ACDC (Advancing Currents Determination in the Solar Chromosphere), working on non-LTE effects; two first-author A&A papers in 2026.
  • Scaled GNN synthesis to 3D atmospheres (millions of nodes): ~10⁶× faster on one H100 GPU than the numerical solver on a CPU cluster — Accelerating 3D non-LTE synthesis with graph neural networks.
  • Joint photosphere–chromosphere sunspot inversions under magnetohydrostatic equilibrium; 6-author collaboration across three institutes.
  • Secured 2M CPU-hours at HLRS (as PI) and 5M at MareNostrum; PI of two more GREGOR campaigns (2025, 2026).
2021 – 2023Co-founder & CTO
UpValuations · startup-valuation platform
  • Co-founded the company with two financial analysts and owned the entire technical side: architecture, development and infrastructure.
  • Implemented the valuation engine in Python (NumPy, pandas, SciPy): discounted cash flow, venture-capital, scorecard, risk-premium and market-multiple methods, with automated PDF reports.
  • Built and ran the production platform, Django/PostgreSQL with Celery/Redis background jobs and online payments on AWS (EC2, S3, RDS) behind nginx and gunicorn, by hand, before AI coding agents existed.
  • Worked directly with clients: 5–10 paying clients and positive EBITDA.
2020 – 2024Doctoral Researcher
Instituto de Astrofísica de Canarias · Universidad de La Laguna
  • Co-developed POLARIS, a large collaborative C++ code for 3D non-LTE spectropolarimetric inversion, and led its statistical-equilibrium solver. It treats level populations and the radiation field as free parameters in one global optimisation.
  • First to apply graph neural networks to non-LTE synthesis and inversion (ApJ 2022); ~10³× faster evaluation than the iterative solver.
  • PI of three observing campaigns (SST 2021, 2022; GREGOR/GRIS 2022) — 22 observing days, full pipeline from proposal to reduced data.
  • Taught two scientific-programming courses at ULL within the FPI fellowship (~30 students a year), 2021 – 2024.
  • ERASMUS+ research stay at the Astronomical Institute of the Czech Academy of Sciences, 2023.
2017 – 2020Research assistant & MSc
Universitat de València · Universidad de La Laguna
  • Zoom-in N-body simulations linking metallicities to star-formation histories (MSc thesis); co-author on MNRAS 2020.
  • Introduction to Astrophysical Research grant: mass estimation of galaxy clusters. Internship at Aras de los Olmos observatory.
  • PRACE Summer of HPC 2020: deep-learning object detection for galaxy orientation.

02Selected results

Inference
3D prominence reconstruction

First fully 3D non-LTE inversion of a solar prominence; resolves the 180° azimuth ambiguity that 1D models cannot.

In prep. (ApJ Letters) · POLARIS
Deep learning
GNN surrogates for physics

Graph networks replace an iterative, non-local solver — two papers: ~10³× faster in 1D, ~10⁶× in 3D.

ApJ 2022 · A&A 2026
Data
38 days, 5 TB of telescope data

Proposal, observation, calibration and analysis of five spectropolarimetric campaigns as PI.

SST · GREGOR

03Education

2024 – 2025Master in Artificial Intelligence (continuing-education master, 60 ECTS), 9.11/10 — Universitat de València 2020 – 2024PhD in Astrophysics, sobresaliente cum laude (highest grade), International Doctorate — IAC & Universidad de La Laguna 2018 – 2020MSc in Computational Astrophysics — Universidad de La Laguna 2014 – 2018BSc in Physics — Universitat de València

04Training

  • NVIDIA Deep Learning Institute — computer vision; multiple data types. 2020
  • PRACE Summer of HPC — 2020
  • DKIST data training — NSO/NSF, 2021 and 2022
  • Deusto Citizen Bootcamp — startup foundation and growth, 2022 – 2023

05Service & talks

  • NeurIPS reviewer, ML4PS workshop, 2022, 2023, 2026
  • Invited: guest lecture, ELLIS Unit Freiburg, 2026 · round table on AI, COEFIS XVI, 2024
  • Talks: SPW10 Kyoto 2022 · EAS Valencia 2022 · COEFIS XV 2023 · SPW11 Prague 2025 · RAS London 2026 · EAS Lausanne 2026
  • Teaching: scientific programming, ULL, 2021 – 2024
  • Supervision: one MSc thesis, one BSc thesis, one internship

06Publications 7 refereed · full list on Work

2020202120222023202420252026

Refereed papers per year, 2020 – 2026; 4 first-author, plus 1 in preparation. Accent bars are first-author.