Arturo Abril Martínez

I recently finished my MSc studies in Artificial Intelligence at Vrije Universiteit Amsterdam and work as a part-time ML engineer at GrwNxt. Previously, I worked as a software engineer at Thales. I hold a BSc degree in Physics from Universidad Complutense de Madrid.

For my master's thesis project, I combined statistics and deep learning to tackle an "inverse" problem in high-energy physics: constraining theory parameters using simulated data. I did this at Nikhef, as part of the ATLAS collaboration. I explored the adaptation of a Transformer architecture equivariant with respect to Lorentz transformations (the symmetry group of relativistic kinematics) to regress likelihood ratios and scores, in a likelihood-free inference setting.

This project sparked my interest in representation learning and geometric deep learning, especially in scalable architectures that use symmetry and geometric structure to work with high-dimensional scientific data. The thesis document is found here.

Mail / Github / LinkedIn

photo of myself

Projects

Advancing the Neural SBI toolkit for SMEFT analyses
We adapt the Lorentz Geometric Algebra Transformer (L-GATr) to perform neural SBI on SMEFT parameters, focusing on likelihood-ratio and score regression tasks. Findings show promising improvements in higher-dimensional (more realistic) analyses.
[Code] [Paper]

Revisiting GraphSAINT with a Feature-Aware sampling method
We reproduce and extend the GraphSAINT sampling method for Graph Convolutional Networks, introducing a feature-biased subgraph sampler and benchmarking F1-micro scores on large-scale datasets (Reddit, Flickr, PPI).
[Code] [Paper] [Original paper]

Convolutional Neural Networks from scratch
Implementing 2D convolutions from scratch, starting from a non-vectorized version (for-loops) to a vectorized version using PyTorch Modules and Functions, working out the backwards by hand, and providing tests against native torch implementations.
[Code] [Paper]

Playing Connect 4 with Monte Carlo Tree Search
We implement a Connect 4 agent using Monte Carlo Tree Search with upper-confidence bound (UCB) to balance exploration and exploitation. The framework generalizes to arbitrary board states and shows improved performance through repeated simulations and root-value averaging.
[Code] [Paper]

Experience

Machine Learning Engineer, GrwNxt
June 2025 - Present

Research Assistant, Nikhef (National Institute for Subatomic Physics)
February 2025 - December 2025

Teaching Assistant, Vrije Universiteit Amsterdam
September 2024 - April 2025

Python developer, Thales
November 2022 - August 2023

Research Assistant (ERASMUS+), GSI Helmholtz Centre for Heavy Ion Research
February 2022 - July 2022