Hybrid & physics-aware ML
Physical laws, constraints and prior knowledge inside learning machines: emulators, parameterizations and hybrid Earth system models that are accurate, consistent and robust.
Full Professor of Electrical Engineering · Head of the Image and Signal Processing group
I develop machine learning to monitor, model and understand our planet. The Earth system is multidimensional, multiscale, nonlinear and connected across many spheres. AI may help to better model the processes, detect changes and extreme events, understand the carbon, water and energy cycles, and discover new laws from observational data and physics understanding. However, prediction is not enough. We need AI that is physically consistent, causal and trustworthy, and that helps us understand a changing climate.
Starting AIREA (Horizon Europe): a RAISE network to develop new AI practices and protocols for agrosystems, 2027–2031.
Starting EarthGenerator (Horizon Europe, GenAI4EU) as PI: generative AI foundation models for Earth science, 2026–2030.
Can AI bridge the gap between the weather and climate? Bridging the weather and climate divide with artificial intelligence, Nature Communications.
The seasonal green wave of vegetation is drifting north-east, and faster every year: Accelerated north–east shift of the global green wave trajectory, PNAS.
A perspective on why machine learning needs epistemology as much as data: AI needs a new philosophy of science, The Innovation.
Joined the Institut Polytechnique de Paris as Visiting Chair Professor (Hi! PARIS, 2026–2029) and the Max Planck Institute for Biogeochemistry, Jena, as Carl Zeiss–Humboldt researcher (2026–2028).
Blaise Pascal Medal lecture at the European Academy of Sciences symposium, CERN, Geneva.
Honoured with the Carl Zeiss – Alexander von Humboldt Research Award, the Blaise Pascal Medal in Earth & Environmental Sciences, and election as AGU Fellow.
“Not only to predict, but to understand: what drives droughts, heatwaves and floods, how ecosystems respond, and where people are most at risk.”
The Earth is a complex, coupled and evolving system, observed today by a wealth of satellite and in situ data. With my team I develop methods that encode knowledge in machines (hybrid and physics-aware machine learning) and decode knowledge from them (explainability, mechinterp, causal inference and equation discovery).
Applications span the biosphere, atmosphere, hydrosphere and anthroposphere, from carbon sinks and crop yields to early warning of food crises and climate-driven displacement, in collaboration with ESA, EUMETSAT, JRC, FAO, UN/WFP and iDMC.
Physical laws, constraints and prior knowledge inside learning machines: emulators, parameterizations and hybrid Earth system models that are accurate, consistent and robust.
Causal inference from time series, explainable AI and symbolic regression to find drivers, mechanisms and governing equations hidden in data and models.
Detecting, forecasting and attributing droughts, heatwaves, floods and wildfires, and tracing their impacts on ecosystems, food security and human displacement.
Satellite data analysis at scale: biophysical retrieval, gap filling, super-resolution, onboard processing and open multimodal foundation models for Earth science.
A PhD in Physics on neural networks and kernel machines to personalise drug dosage for kidney-transplant patients. Few samples, many variables, conflicting goals. Lessons that still shape how I look at data.
Kernel machines for classification, regression, denoising and compression, and their introduction to hyperspectral remote sensing (IEEE TGRS, 2005). Books on kernel methods for remote sensing (Wiley, 2009) and signal processing (Wiley, 2018).
Retrieval, classification, anomaly and change detection from satellites, and the first steps of causal discovery in Earth observation with the ERC Consolidator grant SEDAL.
The 2019 Nature paper with Markus Reichstein and colleagues set an agenda for hybrid modelling in the geosciences, followed by the ERC Synergy grant USMILE and Deep Learning for the Earth Sciences (Wiley, 2021).
Explaining the why of extreme events, early warning of compound climate risk, and open foundation models for a thinking Earth. Towards an AI that is humane, humanist and humanitarian.
A cross-section from first methods to recent perspectives. Everything else, generated from my CV, is on the publications page.
About 12 M€ raised as PI at Universitat de València, with two ERC grants and seven Horizon Europe research actions.
Generative AI foundation models for Earth science (GenAI4EU).
PIEuropean large, open, multimodal foundation models (25 M€ consortium).
PIUnderstanding and modelling the Earth system with machine learning, with V. Eyring, M. Reichstein and P. Gentine.
PIAI for a better representation of processes and extremes in Earth system models.
Copernicus foundation models for a thinking Earth.
A Mediterranean digital twin to understand climate extremes.
European Lighthouse of AI for Sustainability.
AI for agricultural and environmental-pollution science (RAISE network of excellence).
Earlier: SEDAL (ERC Consolidator), XAIDA, DeepCube, iMIRACLI, DeepExtremes, OpenSR, HERMES and Causal4Africa (Microsoft Research). More at isp.uv.es.
Technical Program Chair of IEEE IGARSS 2018 · General Chair of AISTATS 2022 and IEEE MLSP 2012 · Area Chair at NeurIPS and AISTATS · Associate Editor of IEEE Trans. Signal Processing, IEEE Signal Processing Letters and IEEE GRSL · EUMETSAT MTG-IRS Mission Advisory Group.
Essays on AI, science and philosophy, and small existential stories, in Catalan, Spanish, Italian, French and English.
Essays, press & talks →Image and word, compressed to essence. Photographs paired with short verses in Catalan and French.
Open the gallery →Improvisation over structure. Jazz shapes how I think about research, teams and hybrid AI.
Listen along →Reading the game, deciding under uncertainty, playing as a team. The parallels with running a research group are endless.
Courtside notes →