Advancing Discovery, Enhancing Trustworthiness, and Reshaping Scientific Practices
by the guest editors Edina Nemeth (SZTAKI) and Alexandre Termier (University of Rennes – Inria/IRISA, France)
Artificial Intelligence (AI) is rapidly transforming the way science is conducted. From accelerating the discovery of new materials to modeling complex climate systems and supporting biomedical research, AI has become an essential tool for advancing knowledge. By enabling more efficient data analysis, powerful simulations, and new forms of hypothesis generation, AI helps researchers tackle problems that were previously too complex or time-consuming to solve.
In line with the European vision of trustworthy, human-centred, and sustainable AI, this special theme explores the growing role of AI as a catalyst for scientific progress across disciplines. It aims to showcase innovative methods, interdisciplinary collaborations, and real-world applications where AI enhances responsible scientific research and discovery. Contributions highlight both fundamental advances in AI technologies and their application to pressing scientific challenges.
AI as a catalyst for the research process
The articles selected in this section focus on AI systems that directly support the two ends of the research process: the selection of research topics and the evaluation of research through peer review.
In their article Enhancing Reviewer Idenfication with AI, Berndt et al. present an AI-based reviewer identification system that automates the matching of manuscripts to suitable experts by leveraging semantic analysis of abstracts and publication histories. The approach helps ease bottlenecks in peer review and improving the quality, efficiency and fairness of scientific evaluation.
From a different perspective, the article AI Assistant for Research Topic Selection in Higher Education, Stampfl et al. illustrates how large language models can be embedded in higher education as Socratic tutors, guiding students and early-career researchers from vague ideas to consistent, methodologically sound research proposals while demanding conceptual clarity and coherence.
Together, these contributions show AI being used to strengthen the research cycle end-to-end, from early topic design to the assurance of quality in publication.
Building trustworthy AI for scientific decision support
With AI now driving high-stakes decisions, trustworthiness is no longer optional — it’s a core scientific priority.
McAleer et al., in Developing Human-Centred Trustworthy AI as Infrastructure for Reliable Decision Support, report on THEMIS 5.0, which designed methods for assessing accuracy, robustness, and fairness with domain experts in healthcare, maritime operations, and journalism.
Beecks et al., in A Human-Centred AI Approach to Data-Driven Scientific Discovery, present a human-centred data science framework based on Gaussian process models which enable the extraction of interpretable, uncertainty-aware insights while keeping the data scientist in control of the discovery process.
The contribution by Sapidis et al. introduces SemanticRAG: Traceable Answers from Documents and Knowledge Graphs, which combines retrieval-augmented generation with knowledge graphs and explicit provenance, enabling question answering in which every claim is traceable back to specific document snippets or graph triples.
In the article Foundation Models and Trustworthy AI for Environmental Systems, Hatzivasilis et al. examine foundation models through the lens of security assurance and continuous risk assessment, showing how trustworthy AI and security-by-design can be integrated into large-scale monitoring and control platforms.
Across these contributions, trustworthiness is treated as a measurable, context-dependent property that must be engineered, monitored, and enhanced.
Re-engineering the building blocks of AI
A series of articles examine the computational substrates and resource consumption of AI itself.
Biological Reservoir Computing experiments described by Ciampi et al., demonstrate that living neuronal cultures, interfaced through high-density multi-electrode arrays, can act as physical reservoirs for pattern recognition, pointing toward bio-hybrid neuromorphic architectures that combine rich dynamics with potential gains in energy efficiency and adaptability.
Barbierato et al., in Blockchain Energy Costs for AI-Driven Scientific Infrastructure, quantify the energy costs and trade-offs of blockchain-based infrastructures that are often proposed as trustworthy backbones for AI-driven scientific platforms, highlighting that verifiability and decentralisation have a tangible environmental price.
Complementing this in a second contribution, Barbieto et al. argue that in scientific contexts Accuracy is not Enough, and that computational efficiency and alignment with domain knowledge remain key design objectives when deploying AI at scale to ensure reproducibility and value in gaining additional insights.
Together, these works emphasise that advancing AI for science also requires rethinking its basic components to make them more sustainable, explainable, and physically grounded.
AI for imaging, sensing, and scientific observation
Many scientific advances depend on extracting meaning from visual and spatial data, an area where AI is rapidly becoming indispensable.
Szentirmai et al. discuss Lightweight on Device AI showing how resource-aware models running on augmented reality devices can act as flexible scientific infrastructure, supporting universally designed experiments and studies in the field.
Computational Imaging research presented by van Leeuwen et al. explores how AI can push beyond traditional image reconstruction pipelines, enabling new modalities, reducing computational cost, and improving image quality in resource-constrained settings.
The article presented by Ćeranić et al. describes a hybrid method for Detecting Small Changes in 3D Aerial Scans, including structural or terrain changes of real urban environments.
Burnet and Parisot report on Fully Automated Detection of Harmful Cyanobacteria Blooms in Lakes Using Photo Traps and YOLO-based Object Detection, enabling near-real-time early warning in critical water reservoirs.
These contributions exemplify how AI expands our observational capabilities, turning heterogeneous, high-volume image streams into operational scientific insight.
AI for society and for modelling complex systems
Finally, the special theme highlights AI’s role in modelling complex systems and supporting societal missions.
An AI-enhanced Operational Picture for Public Safety Operations combines heterogeneous textual sources with knowledge graphs to build semantically coherent situational overviews for emergency responders, addressing both information overload and domain heterogeneity is presented by Pöschl et al..
Segura Ortiz et al. show how Context-guided Evolutionary Algorithms are transforming gene regulation inference by integrating multiple computational approaches with biological knowledge.
In energy and climate-related modelling, a generative adversarial surrogate trained on high-fidelity simulations learns to reproduce wind-turbine wakes at hub height, offering orders-of-magnitude speed-ups that enable optimisation and control studies that would otherwise be computationally prohibitive.
Work on deep learning-enhanced multi-scale simulations of molecular systems shows how AI can bridge modelling scales, coupling atomistic detail with mesoscopic or continuum descriptions in ways that preserve essential physics while reducing computational burden.
Across these efforts, AI serves as a mediator between data and theory, enabling models that are fast enough, and expressive enough, to be embedded into real-world decision processes.
Concluding remarks
Taken together, the articles in this special theme make clear that AI technologies are no longer peripheral tools in scientific research. AI for Science represents an evolving ecosystem of tools, infrastructures, and practices that cut across disciplines. They demonstrate AI helping to organise the scientific process, deepening trust in model-based decisions, facilitating the understanding of complex systems, extending our senses, and enabling new forms of modelling and control. At the same time, this transformation raises fundamental questions about trust, transparency, energy consumption, and the evolving role of human expertise in the scientific process.
We hope this collection will not only inform readers about current advances, but also inspire new collaborations at the interface of AI and the many scientific disciplines it now helps to transform.
As we witness the rapid development of AI technologies themselves, we are also witnessing their growth into a fundamental scientific infrastructure underpinning discovery across disciplines. Emergent approaches such as Denario or Sakana [L1] even aim to automate key steps in scientific discovery — from hypothesis generation to experimental design. This suggests that we may be only at the beginning of a deeper transformation in how scientific knowledge is generated.
Link:
[1] https://github.com/AstroPilot-AI/Denario
Please contact:
Edina Nemeth, SZTAKI, Hungary
Alexandre Termier
University of Rennes – Inria/IRISA, France
