The EPIA 2026 International Conference on Artificial Intelligence

Amparo Alonso Betanzos

University of Coruña

Frugal AI: Sense, Sustainability and Technological Sovereignty

Professor at the Universidade da Coruña and coordinator of the LIDIA Laboratory (CITIC); visiting professor at NTNU (Norway). Her research focuses on sustainable and responsible AI. She has served as President of the Spanish Association for Artificial Intelligence (AEPIA); she is a Senior Member of IEEE and ACM, and a Member of the Royal Academy of Sciences. She was a member of the AI Advisory Council (CAIA) of the Spanish Ministry for Digital Transformation, and contributed to the drafting of Spain's AI R&D&I Strategy and the HispanIA 2040 Report of the National Office for Foresight and Strategy. She currently serves on the Spanish Committee on Research Ethics and the AI Ideas Laboratory of AESIA.

Abstract

The success of Artificial Intelligence has rested on an implicit premise: more data, larger models, and greater computing power yield better results. This logic has worked, but it carries a cost that rarely makes the headlines. Training a state-of-the-art model can emit thousands of tonnes of CO₂ and consume hundreds of thousands of litres of water. By 2030, AI is estimated to account for more than 30% of global energy consumption — a trajectory that, left uncorrected, is unsustainable environmentally, economically, and geopolitically. Frugal AI emerges as a response to this challenge: an approach that seeks to achieve the best possible performance at the lowest resource cost, both during training and inference. The goal is not to sacrifice quality, but to rethink how it is achieved. The strategies we will explore include improving training data quality, developing more energy-efficient architectures, and learning techniques that enable useful models to be built from fewer examples and with lower energy consumption. Beyond environmental sustainability, frugal AI has strategic implications. It reduces dependence on massive computing infrastructure, opens the field to institutions and countries with fewer resources, and favours systems that are more transparent and interpretable. In this sense, it aligns with European priorities for developing AI that is safe, sovereign, and responsible.

Francisco C. Pereira

Technical University of Denmark

The Unknowable Future: AI, Simulation Surrogates, and Deep Uncertainty

Francisco Camara Pereira is a Professor at DTU, where he leads the Intelligent Transport Systems group. His research focuses on the methodological combination of Machine Learning and Transport Research, addressing challenges such as climate adaptation, demand modeling, traffic prediction, data collection, simulation metamodeling, and anomaly detection. He has been a Marie Curie fellow twice (2011 and 2016) and is currently a Novo Nordisk Data Science Distinguished Investigator. He has published over 80 articles in both Machine Learning and Transport Research. Before joining DTU, he was a Senior Research Scientist at SMART/MIT (2011-2015) and an Assistant Professor at the University of Coimbra (2005-2015).

Abstract

What climate will a metro line opened today face in 2090? No scenario model can tell you — not because the science is bad, but because across a sixty-year horizon, the uncertainties compound past the point where any one path is meaningful. We build anyway. We have to. So the question shifts: short of prophecy, what do you actually anchor a fifty-year decision to? This talk explores how reinforcement learning and AI-based simulation surrogates can support decisions about the future of urban transport in the face of precisely this kind of uncertainty. Working within the MAAT project, we have developed a framework that combines climate projections, hydrological flood modelling, and transport network simulation to identify where and when adaptation measures should be deployed in flood-prone cities — without assuming we know which future will actually arrive. A central methodological challenge is computational: coupling RL with high-fidelity urban flood simulation at the scale needed for policy-relevant results is prohibitively expensive. We address this through APEX, a surrogate modelling approach that reduces simulation time from days to minutes while preserving the problem's structure. We present results from Copenhagen, and reflect on what they suggest about the possibility of planning well, not by predicting the future, but by learning to act wisely in spite of not knowing it.

Michael Schroeder

TU Dresden

From AlphaFold to Interaction Embeddings: AI for Structure-based Drug Repositioning

Michael Schroeder is a professor in bioinformatics at Dresden University of Technology, TUD, where he teaches bioinformatics and AI. His research is dedicated to AI for improved diagnosis and treatment of cancer and infectious diseases. He has published widely on these topics with an H-index of over 50. He served as director of his institute and has co-founded three AI-based spin-off companies from TU Dresden. He has graduated over 20 PhD students and10 former group members are professors or group leaders now. He did a PhD in non-monotonic reasoning at Hanover University and FCT NOVA.

Abstract

AlphaFold transformed structural biology by demonstrating that deep learning can predict protein structures at proteome scale, increasing the number of available structures from around one hundred thousand experimentally determined models to more than two hundred million predicted ones. Yet protein structure alone is only part of the story. Biological function ultimately emerges from molecular interactions. In this talk I will argue that the next challenge is not simply predicting structures but learning useful representations of molecular interactions. Instead of comparing proteins or drugs individually, we represent protein–ligand interactions as standardized fingerprints that capture the underlying binding geometry and chemistry. These interaction embeddings make it possible to search large structural databases for analogous binding mechanisms across otherwise unrelated proteins. Using examples ranging from malaria drugs in cancer to cancer drugs in autoimmune disease and pain medication in infectious disease, I will illustrate how similar interaction patterns can reveal unexpected therapeutic opportunities. The examples demonstrate how representation learning enables transfer across distant protein families that share little sequence or structural similarity. Finally, I will discuss how AlphaFold changes the landscape for AI-based drug discovery. While AlphaFold provides an unprecedented structural foundation, its models largely lack bound ligands and interaction information. This raises new AI challenges: how can interactions be inferred, represented, and indexed at AlphaFold scale? Can interaction embeddings become the analogue of sequence embeddings for molecular recognition? These questions point toward the next generation of AI methods for molecular discovery.

EPIA 2026

Invited Speakers

Amparo Alonso Betanzos

Francisco C. Pereira

Michael Schroeder

Invited Speaker

João Pavão Martins

Sponsors & Partners