Scientific publications
Computational mechanics, optimal control, energy and materials simulation, plus the data and code that stayed open. Each entry carries the reference exactly as published and one line on why it matters.
Where there is a DOI or an official record, it is linked; where I could not confirm one, the entry has no link and says so. I would rather leave a link out than guess an identifier. The full, up-to-date list is on my ORCID.
Scientific publications
Dimat Multiscale Simulation Toolkits For Industrial Modelling
Contribution. How to fit atomistics (LAMMPS), finite elements (CalculiX) and trained models into a single chain, so that process simulation — polymer extrusion, yarn formation — stays connected to material property prediction. The point is traceability: each level still rests on the one below.
Multiscale Modeling of Textile Yarns and Nets via Particle-Based Equilibrium and Finite Element Analysis
Contribution. A braided yarn or a net has so many contacts that a continuum model stalls before it starts. Here the filaments are first solved as chains of particles with stiffness and friction until they reach equilibrium, and that geometry — stress state included — is handed over to finite elements for loading. The hard part, and what the paper is about, is not losing that equilibrium in the handover.
Advancing Polymeric Yarn Manufacturing: A Hybrid Modelling Approach Combining AI and Molecular Dynamics
Contribution. Molecular dynamics and trained models combined to predict the behaviour of a polymeric yarn without rerunning the atomistic simulation at every point of the design space.
Optimizing AI-Enhanced Neural Network Subroutines for Plasticity in FEM
Contribution. A material subroutine that calls a neural network at every integration point is easy to write and hard to make fast enough. The work proposes and validates a methodology for optimising such an ML-UMAT against a specific plastic potential — the only way to know whether the network is approximating what it should.
Energy Management System for a Residential Positive Energy District Based on Fuzzy Logic Approach (RESTORATIVE)
Contribution. A 360-home district in Bilbao with central storage, where fuzzy logic decides when to charge and discharge instead of a fixed threshold. It reaches up to 75.6 % self-sufficiency with 6–12 year payback — and, which is the part usually missing, it measures what happens when prices or climate move outside the expected range.
Digital Tools for Modeling and Simulation of Glass-Forming Process
Contribution. Glass forming, by mould and by laser, modelled end to end with free software: parametrised geometry in FreeCAD, solving in CalculiX, and the whole workflow automated in Python, STEP-to-FRD conversion included. The argument is not numerical but about access: a small company cannot pay yearly simulation licences.
Multilevel Selective Harmonic Modulation via Optimal Control
Contribution. Picking the switching angles of a multilevel converter to cancel specific harmonics is normally solved as a system of nonlinear equations, with the number of switches fixed beforehand. Here the problem is posed as optimal control with a piecewise affine penalisation: Pontryagin’s maximum principle forces the solution into staircase form, and how many switches there are and where they fall stop being an input and become part of the answer. The solution is also unique and continuous with respect to the target frequencies.
Integration of graphene into nanoelectronic devices: insights from atomistic simulations
Contribution. The first one. From the electronic-structure years at SIMUNE Atomistics, working with SIESTA: the starting point of everything that came after.
Open data and software
Climate and Crop variables of tomato greenhouse simulation
Contribution. The full time series of a simulated tomato greenhouse: outdoor climate, indoor climate, carbon allocation across plant parts, leaf area index, water uptake and window and screen openings — about 33 variables. From the greenhouse modelling years at the University of Deusto, published openly so that a controller can be trained or checked without building the greenhouse.
DyCon Toolbox
Contribution. A MATLAB toolbox for nonlinear optimal control problems, with derivatives by automatic differentiation (CasADi) and both routes available: the direct method — discretise state and control and solve the resulting nonlinear program with IPOPT — and the indirect one — integrate the adjoint backwards and descend along the gradient. The idea was that a control problem should fit in a few readable lines and that someone else should be able to repeat it.
Navindoor
Contribution. A free MATLAB library for simulating indoor positioning systems end to end: floor plans, trajectories and walker dynamics, synthetic signal generation — received signal strength, time and angle of arrival, barometer, magnetometer, inertial — and algorithm comparison, with both a GUI and a CLI. It lets you evaluate a localisation algorithm before installing a single beacon.
Academic work
Reinforcement learning applied to the shepherding problem
Aprendizaje por refuerzo aplicado al problema del pastoreo
Contribution. A dog that can only repel has to drive a flock to a target. The work builds the dog–sheep interaction model, defines a heuristic “push and turn” policy as the reference, and then trains a DDPG policy pre-trained on that heuristic. Reference first, learning second: the same order I still work in.
Optimising numerical basis sets on dimers. Development, improvements and documentation of the SIESTA code
Optimización de bases numéricas en dímeros. Desarrollo, mejoras y documentación del código SIESTA
Contribution. In plane-wave DFT the basis quality is one number and the energy falls monotonically with it: raising it trades cost for accuracy and little else. In SIESTA, which uses confined numerical atomic orbitals, the basis is set by several parameters — cut-off radii, the second split-valence radius, and the λ and Z of the Yukawa potential — that behave non-monotonically and admit several equivalent sets of values. Which is why most people reach for a known-good basis instead. The work optimises those parameters on the cheapest system there is, the dimer, and then checks whether the result transfers to a different chemistry: on the water dimer the basis reaches meV accuracy and matches bases optimised specifically for liquid water and ice (238 meV binding energy against their 236, and 228 with plane waves); on crystals it reproduces the plane-wave converged lattice constant of bcc Li and beats SIESTA’s own default basis, which is larger. This is where not trusting a calculation until it is checked against a reference started.