Ibón García Gómez | Physics & Machine Learning

I develop physics-informed deep learning methods for precision instrumentation and astrophysical surveys.

My work sits at the intersection of experimental physics and machine learning. I design neural architectures that exploit the underlying symmetries and spectral properties of physical systems — from sub-micron position reconstruction in silicon particle detectors to automated detection of stellar streams for the ESA ARRAKIHS mission.

I am currently completing my MSc in Particle Physics & Cosmology at the Universidad de Cantabria, and actively seeking PhD positions starting Fall 2026 in areas involving ML-driven detector development, computational astrophysics, or instrument science.

Email GitHub Kaggle LinkedIn

Universidad de Cantabria

Santander, Spain

Research Projects

Selected Work

ESA ARRAKIHS Collaboration Master's Thesis

Deep U-Net for Faint Stellar Stream Detection amidst Galactic Cirrus

Detecting stellar streams — fossil remnants of tidally disrupted dwarf galaxies — is a key challenge in Near-Field Cosmology. These structures are extremely faint and frequently overlap with galactic cirrus (interstellar dust), which has similar morphology but distinct spectral signatures. I designed a complete pipeline: a physics-informed synthetic data generator using \(1/f^\beta\) fractal noise injected into real SDSS/WISE backgrounds (20,000 images), a Depth-4 U-Net with a Dice-weighted hybrid loss optimized via Bayesian HPO (Optuna), and 4× Test-Time Augmentation for robust inference. The model exploits multi-band spectral decomposition (g/r/W1/W2) to disentangle stellar emission from dust contamination.

> 0.84

Stream IoU

> 93%

Stream Recall

20,000

Synthetic Samples

GitHub Repository → Kaggle Notebooks →

TensorFlow U-Net Optuna AstroPy FITS/HiPS Mixed Precision

Detector Instrumentation Beats 2024 JINST SOTA

Sub-Micron Position Reconstruction in 3D Silicon Detectors via Multi-Domain CNNs

High-precision hit reconstruction in 3D-trench silicon pixels is limited by intra-pixel ambiguities arising from the detector's internal symmetry. I developed a dual-branch neural network (1D-CNN + physics-informed MLP) that fuses raw TPA-TCT waveforms with a 110-dimensional multi-domain feature vector built using the pixel's \(C_{2v}\) point-group symmetry. Group-theoretic irreducible representations (Irreps) provide a dedicated basis for resolving quadrant sign ambiguities. The system was optimized via an 11-hour Bayesian search over 14 hyperparameters and validated with multi-seed statistical analysis.

0.80 µm

Euclidean MAE

91.1%

Quadrant Accuracy

1.5 → 0.8 µm

vs. JINST 2024

GitHub Repository → Kaggle Notebooks → Dataset →

PyTorch 1D-CNN MLP Optuna Group Theory Signal Processing

Gravitational Wave Physics Bachelor's Thesis

Gravitational Wave Emission from R-Mode Instabilities in Neutron Stars

R-mode oscillations in rapidly rotating neutron stars are driven unstable by gravitational radiation reaction (CFS instability) and represent a promising continuous gravitational wave source for LIGO/Virgo. I built a comprehensive numerical simulator (~2,300 lines) that solves the coupled evolution equations for the r-mode amplitude \(\alpha(t)\) and angular velocity \(\Omega(t)\) following Owen et al. (1998) and Lindblom et al. (1999). The tool includes viscous damping, non-linear saturation, thermal evolution with modified URCA cooling, and a graphical interface for visualizing spacetime deformation with LIGO sensitivity overlays.

GitHub Repository →

Python SciPy (RK45) Matplotlib ODE Integration Tkinter GUI

What I'm Working On

Current Activity

Sub-Pixel Resolution in Hexagonal Detector Arrays

Developing MLP/CNN architectures for position reconstruction in hexagonal pixel geometries. Active research project — Universidad de Cantabria.

Stanford CS224W — Graph Neural Networks

Self-directed study of GNN architectures (GCN, GAT, GraphSAGE). Coursework to be published on GitHub upon completion.

Santander X Spain Awards — Precision Dosimetry in Theranostics

Exploring neural network approaches for precision dosimetry with Lutetium-177 in nuclear medicine. Technology & Science category.

Research Directions

Open Questions I Want to Pursue

01 Can multi-band spectral decomposition reliably separate stellar streams from galactic cirrus at the detection limit, without requiring spectroscopic follow-up?
02 What is the fundamental spatial resolution limit of position-sensitive silicon detectors when exploiting physics-informed neural architectures?
03 How can graph neural networks leverage the irregular topology of non-Cartesian detector geometries (hexagonal, 3D-trench) for reconstruction tasks?

Background

Education

2025–26

MSc in Particle Physics and Cosmology

Universidad de Cantabria · Expected July 2026

Thesis: Deep learning for Low-Surface Brightness structure detection (ESA ARRAKIHS collaboration)

2021–25

BSc in Physics

Universidad de Cantabria

Thesis: Gravitational wave emission from r-mode instabilities in neutron stars

2024–25

University Micro-Credentials in Data Science & AI

Universidad de Cantabria — Data-Oriented Programming and Artificial Intelligence in Engineering (I & II)

2026

CS224W: Machine Learning with Graphs

Stanford University (self-directed) · In progress

Technical Profile

Skills & Tools

Machine Learning

TensorFlow/Keras, PyTorch, Optuna (Bayesian HPO), Semantic Segmentation (U-Net), 1D/2D CNNs, MLPs, Mixed Precision Training, Test-Time Augmentation, Custom Loss Functions

Scientific Computing

NumPy, SciPy, Pandas, Scikit-learn, Matplotlib, Seaborn, AstroPy, FFT/Signal Processing, ODE Integration (RK45), Monte Carlo Methods

Physics Domains

Particle Detector Instrumentation (TPA-TCT, Silicon Pixels), LSB Astrophysics (Stellar Streams, Galactic Cirrus), Gravitational Wave Physics (R-modes, LIGO), Group Theory (C₂v)

Engineering

Python, Git/GitHub, Kaggle, FITS/HiPS (Virtual Observatory), LaTeX, Linux, Jupyter, Concurrent Programming, GUI Development (Tkinter)