Andrés Vicente Arévalo
I turn noisy measurements into physical answers — and make the slow part fast.
Solar physicist at the Leibniz Institute for Solar Physics. I build inversion codes that recover the Sun's magnetic field from polarised light, and graph-neural-network surrogates that make the radiative transfer behind them 10³ to 10⁶ times faster.
Now looking to bring that toolkit — inverse problems, 3D modelling, PyTorch and supercomputer-scale data — to research-scientist, data-science and machine-learning roles.
Selected figures
Three results, one visual each
All three are live: rotate the prominence, pick any pixel of the sunspot, turn the graph.
Two dense feet bridged by rarefied plasma; depth along the line of sight recovered from radiative transfer alone. Orange is the ortho-helium density from the POLARIS inversion, green the field lines of the inferred magnetic field. In preparation (ApJ Letters).
The FIRTEZ inversion of a sunspot observed with CRISP at the Swedish Solar Telescope in four lines (Fe I, Na I D1, Ca II and Mg I b2), with non-LTE and magnetohydrostatic equilibrium, on a geometrical height scale. Pick a pixel on either map for its observed and fitted Stokes profiles and its temperature, velocity and magnetic field against optical depth; the field is in the observer’s frame, z along the line of sight. The first pixel sits in an umbral flash: the Ca II core is in emission and the chromosphere above it is hot. A&A 2026.
1D on one CPU; 3D on one H100 GPU against a CPU cluster.
A graph network learns the iterative, non-local synthesis once and evaluates it cheaply thereafter: 10³× faster in 1D, 10⁶× in 3D. The graph is the paper's schematic of the one the 3D model works on: the populations are inferred in the target column (red) from the context points linked to it (blue). The model's own graphs have 11 × 11 × 55 nodes. ApJ 2022 · A&A 2026.
From photons to inference
The pipeline I build, end to end
PI on SST and GREGOR runs; calibration and reduction of spectropolarimetric data.
Four Stokes profiles per pixel: the raw signal that encodes the magnetic field.
Iterative, non-local forward model. Slow — which is the problem to solve.
A graph network learns stage 3 once, then evaluates it 10³× faster in 1D and 10⁶× in 3D.
Global minimisation over 3D geometry; error bars, no 180° ambiguity.
Every stage is one I have written code for — from telescope calibration scripts to the GPU-trained surrogate that replaces stage 3.
Timeline
2014 → today, hover to read
Why industry, why now
The problems I like are the ones where the model is expensive and the data is messy.
That describes a solar prominence. It also describes demand forecasting, sensor fusion, risk models and simulation-based inference. Ten years of squeezing physical answers out of imperfect measurements — with error bars — is the part I want to keep. The subject can change.