Gravity & holography
Strongly coupled quantum matter, mapped onto gravity in one more dimension. Numerical relativity in asymptotically AdS spacetimes follows it in real time, and quasinormal modes of black branes probe its stability.
PhD Researcher in Theoretical Physics
Institute of Cosmos Sciences (ICCUB) · Universitat de Barcelona
I am a PhD student in theoretical physics at the Institute of Cosmos Sciences (ICCUB), Universitat de Barcelona, supervised by David Mateos and Raúl Jiménez and funded by an FPI fellowship (Ministerio de Ciencia e Innovación) as part of the project Dark Energy and the Origin of the Universe. I expect to defend my thesis in 2027 and am looking for postdoctoral positions starting in autumn 2027.
I work on strongly coupled matter and gravity, and on what they tell us about cosmology and astrophysics. Much of it revolves around first-order phase transitions, in the early universe or inside neutron star mergers. These proceed through bubbles that nucleate, expand and collide, and the fluid motion they stir up is a source of gravitational waves. I study how these bubbles expand when the equation of state is realistic rather than conformal, and I use holography (AdS/CFT) to describe the strongly coupled, far-from-equilibrium dynamics that sits underneath.
Holography also turns many physical questions into inverse problems. One example is finding the gravitational dual of a given gauge theory using only its equation of state. Problems like this are stiff and highly non-linear, and I tackle them with physics-informed neural networks (PINNs). That work has fed back into machine learning itself: new regularization schemes that let PINNs transfer-learn into stiff regimes, and the question of what a network actually learns when it solves a whole family of differential equations at once.
I am a member of the Simons Initiative on the Geometry of Flows, where I work on turbulence, and I have contributed to Denario, a multi-agent AI system for scientific research. Most of my work is collaborative. A research stay with Pavlos Protopapas at Harvard SEAS in 2024 started a series of joint papers, and in December 2026 I will visit the Institute for Advanced Study in Princeton to work on turbulence with James Stone.
Before my PhD I completed an MSc in Theoretical Physics at the Instituto de Física Teórica (IFT), UAM-CSIC — where my thesis on Quantum Chaos in de Sitter Space and Centaur Geometries was supervised by Juan F. Pedraza and Ayan Kumar Patra — and a BSc in Physics at the Universidad Complutense de Madrid.
Click a region to read about it.Tap a region to read about it.
Strongly coupled quantum matter, mapped onto gravity in one more dimension. Numerical relativity in asymptotically AdS spacetimes follows it in real time, and quasinormal modes of black branes probe its stability.
A strongly coupled fluid becomes a black-brane geometry: bubble walls and turbulent eddies turn into dynamics of spacetime, giving access to quantities that are hard to compute from the fluid alone. Holography also tests effective scalar–fluid models of phase transitions, and gives the friction coefficient that sets how fast bubble walls move.
Bubbles in first-order phase transitions and the gravitational waves their flows source. Non-conformal equations of state obstruct some expansion modes and, for QCD-like theories, rule out detonations. Also turbulence, in the Simons Initiative on the Geometry of Flows.
The gravitational dual of a gauge theory, reconstructed from its equation of state alone by solving the bulk equations as an inverse problem with PINNs, now including large hierarchies, false vacua and skipping RG flows.
Physics-informed neural networks for stiff and inverse problems, such as the Calderón problem of impedance tomography, and for Hamilton–Jacobi–Isaacs equations, where a small residual can still hide the wrong solution. Also Denario, a multi-agent AI system that proposes ideas, writes code and drafts papers.
Multi-head networks learn whole families of solutions; regularizing their latent space lets them transfer-learn into stiff regimes. For the Burgers, heat and wave equations that space reduces to a few components that are reproducible across training runs.
The common thread: matter that is strongly coupled and far from equilibrium, from the early universe to neutron star mergers. Holography turns it into gravity, hydrodynamics describes how it flows, and machine learning solves the hard problems in between.
An improved effective description of thermal first-order phase transitions in relativistic theories, in which a scalar order parameter is coupled to a fluid, tested against a microscopic calculation in a holographic model. The sign of the specific heat decides the spinodal instability: non-hydrodynamic when it is positive, an unstable sound mode when it is negative. It also identifies a friction coefficient that sets the bubble-wall velocity and can be extracted from a holographic correlator.
For Hamilton–Jacobi–Isaacs equations a small PDE residual does not certify a learned solution: standard PINNs converge to almost-everywhere solutions that differ from the viscosity solution by order one. Giving the network geodesic-distance information recovers the viscosity solution, and a mesh-free semiconcavity test detects the failure without knowing the exact answer.
Self-similar bubble solutions for equations of state ranging from the bag model to QCD-like theories, with pure SU(3) Yang–Mills as a benchmark. Non-conformality creates new obstructions to expansion and a new branch of “shocked detonations”. For QCD-like theories every detonation is excluded, which lowers the kinetic energy available for gravitational waves.
A multi-head PINN learns a latent manifold for a whole family of PDE solutions. For Burgers dynamics, two to four principal components carry about 95% of the latent variance. How they spread over Fourier shells is the same in independent training runs, so it gives an observable of the solution space itself. Heat and wave equations serve as checks.
Extends the holographic reconstruction from thermodynamic data to theories with large hierarchies of scales, false vacua and skipping RG flows. Scalar potentials are recovered deep into the false-vacuum regime, including regions the input data do not probe directly.
Neural inversion of the Calderón problem of electrical impedance tomography from finite boundary data, using randomized multiscale wavelet excitations and Fourier-feature encodings. The dominant conductivity structure is recovered to within roughly 3–12%.
Introduces the latent-space analysis for the viscous Burgers equation: a network trained on many initial conditions organizes its latent space onto a low-dimensional solution manifold.
A modular multi-agent AI system that assists across the research workflow, from idea generation and literature review to writing and running code, analysis and drafting papers. It is demonstrated on problems from many scientific disciplines.
Multi-head PINNs learn the space of solutions of a family of equations instead of a single solution. A unimodular regularization of that latent space makes transfer learning much more efficient, reaching stiff, multiscale and inverse problems that are otherwise hard to solve.
Treats holography as an inverse problem: the five-dimensional gravitational dual of a strongly coupled gauge theory is reconstructed from its equation of state alone, using physics-informed neural networks.
Citation counts: the higher of Google Scholar and INSPIRE-HEP for each paper, November 2026. Full list and citation tracking: Google Scholar · INSPIRE-HEP · ORCID · arXiv.
The machine-learning module of the Mathematical and Statistical Techniques course in the Master's in Astrophysics, Particle Physics and Cosmology at the University of Barcelona. The lectures introduce neural networks as a numerical tool for physics, with a hands-on focus on physics-informed neural networks (PINNs) for solving differential equations, and an introduction to dimensionality reduction. All notebooks, examples, and exercises are openly available on GitHub.
Interested in the lectures or in using the materials? The notebooks are self-contained and free to use — feel free to reach out with questions or feedback.
PhD supervision
Education
Research experience
Fellowships & awards
Talks & posters
Schools, workshops & research stays
Teaching
Service & memberships
Open-source code
Media & outreach
Skills & tools
Beyond research
Amateur astronomy (telescope operation and calibration), piano (professional-level training at the Conservatorio Profesional de Música Jacinto Guerrero), reading and music.
Happy to discuss research and collaborations in holography, phase transitions and gravitational waves, turbulence, or machine learning for physics.