Research highlights

Hospitals hold critical MS data but cannot share it due to privacy constraints. FAIR researchers built an ecosystem where the resulting models are shared instead of patient data, enabling collaboration across institutions without risk.

Federated and centralized MS-progression models match international state-of-the-art benchmarks. Tools include MRI super-resolution, schema matching, federated learning infrastructure and explainable models. A key clinical finding: progression history is a stronger predictor of future disability than relapse or treatment history.

Modern machines—such as heating systems, production equipment or vehicle components—can behave very differently depending on how they are used. This makes it difficult to detect early signs of wear or to automatically adjust control settings when the environment changes. This use case explores how AI can make machine fleets more adaptive, self‑monitoring, and self‑optimizing.

By combining real‑time condition monitoring with intelligent control, machines can learn from each other, adapt to new workloads, and operate more efficiently. Practical applications include early detection of mechanical issues, safer and more reliable operation, and energy‑efficient control in systems such as heating networks.

HEAT.NET is an integrated monitoring-and-control system that brings smart features together on one platform, combining at least three heat-supply setups and supporting both open-loop and closed-loop control. The system includes multiple moving components—such as gears and belts—that can be tracked using sensory-AI condition monitoring. Reinforcement learning for closed-loop control is validated on extensive simulation data and supported by limited real-world measurements. Context-aware gear monitoring allows gear conditions to be assessed under different operating regimes. Development builds on toolboxes such as Edge AI4SM for on-device monitoring and RL4CTRL for learning-based control.

Understanding how individual cells behave inside a tissue is essential for diagnosing diseases, improving treatments, and studying how plants and animals grow. This requires analysing large, detailed datasets that describe which genes each cell activates (single-cell data) and how cells interact with their neighbours (spatial omics data.)

SpatialData is a universal data format for organizing spatial omics information and supports large-scale AI experimentation. Building on this foundation, SPArroW is a toolbox for spatial transcriptomics (gene-activity measurements with spatial context), and Harpy is a matching toolbox for spatial proteomics (protein-level measurements in tissue). Both toolboxes are under journal review and have already been used in studies of wheat roots and melanoma samples. The teams also contributed to improved cell-cell communication models, better uncertainty modeling for single-cell predictions, and helped establish an international benchmarking platform for evaluating single-cell omics methods.

Agrifood environments are unpredictable. Under varying lighting- , crop- , and soil conditions, robots must operate safely and securely. FAIR partners advance field‑ready AI for these conditions by improving sensing, mapping, and farm management to pave the way for robotics to support with the most labor-intensive tasks in horticulture and arable farming.

The Light-Invariant Spectral Autoencoder is a new model for estimating grape characteristics — Brix, acidity, and weight — from hyperspectral imaging under variable lighting conditions. A new Treebot sensor module integrates camera-LiDAR for autonomous data collection, enabling future scaling of sensor-fusion and reinforcement-learning control. The ARTOF framework has been upgraded to support validation of RL and vision-language-action models in realistic field settings. Due to its success, the AI technology researched in viticulture is now also scaled to other horticulture cultivations and arable farming applications like weed control.

Movement data from sports or rehabilitation is complex and noisy. FAIR teams built tools that extract meaningful patterns from this data, predicting actions in football and identifying clinically relevant movement motifs, all while processing multi‑sensor signals much faster than before.

A new football event-prediction model improves accuracy by 3.5 percentage points when predicting the next on-field action. A motif (recurring pattern) discovery algorithm achieves 10–20% higher precision and recall when identifying clinically relevant movement patterns in movement data about rehabilitation exercises. Finally, a new algorithm for analyzing multi-channel time series data can learn models about movement data 2–10× faster while maintaining accuracy.

Together these advances increase analytical speed and detection accuracy for sports, rehabilitation and movement science.

Air quality, climate conditions, and environmental risks change quickly, and it is often difficult for cities, citizens, and policymakers to act in time. This use case focuses on AI tools that help us better understand, predict, and respond to these changes. The work improves air‑quality forecasting, strengthens the interpretation of remote‑sensing images for climate adaptation and disaster response use, and speeds up models used to simulate environmental processes.

Decision-focused learning is used to link predictive air-quality models with prescriptive optimization, allowing uncertain environmental parameters to guide downstream decision tasks. Self-supervised representation learning is applied to remote-sensing imagery, where dense and object-centric SSL methods are evaluated as strong building blocks for specialized environmental vision-language models, while their operational value is evaluated through integration into future agentic architectures to support disaster management. Deep-learning-based surrogate modelling is being developed to approximate expensive physics-based simulations, including computational-fluid-dynamics models used in environmental analysis, with a toolbox planned for release in late 2026.

Many real-world decision-making problems, such as proposing mitigation policies for epidemics, require balancing effectiveness with other objectives, including fairness.

FAIR researchers created AI systems that integrate fairness directly into decision-making, enabling policymakers to make effective decisions without disproportionately burdening specific groups.

Multi‑objective reinforcement learning (RL) models can optimize simultaneously for health outcomes and fairness.

A social burden fairness model ensures that epidemic restrictions are balanced across age groups while maintaining comparable hospitalization control.

Example paper:

Reymond, M., Hayes, C. F., Willem, L., Radulescu, R., Abrams , S., Roijers, D. M., Howley, E., Mannion, P., Hens, N., Nowe, A., & Libin, P. Exploring the Pareto front of multi-objective COVID-19 mitigation policies using reinforcement learning. Expert Systems with Applications, 249 (Part C), 2024.

Industrial assets generate large volumes of time‑series data that reflect wear, operating conditions, and emerging failures. FAIR partners have devised methods that turn these long, complex data streams, while creatively combining them with other data sources as log records and technical specifications, into earlier warnings and clearer diagnostic signals, supporting maintenance decisions in settings such as wind turbines, compressors, water pumps and other field‑deployed equipment.

Advanced multi-view data exploration and modelling strategies integrate diverse learning paradigms (neural, probabilistic, symbolic) combining time series with other data modalities, such as log records and tabular data. The strategies are formalized into standardized workflows (recipes) for the effective analysis, integration, forecasting, and extraction of actionable insights from multi-source industrial data, while addressing industry-relevant applications - including anomaly detection, failure prediction, state-of-health estimation, and product usage monitoring - in a methodologically sound and well-structured manner. The proposed workflows have proven to be highly adaptable, accommodating variations in industrial contexts, application requirements, and data availability while providing a systematic and transferable framework for industrial AI applications.

Example papers:

• Murtas, G., Boeva, V., & Tsiporkova, E. (2026) Neuro-LENS: a neuro-symbolic framework integrating incomplete background knowledge and deep learning. Neurosymbolic AI journal.

• Fingerhut, F., & Tsiporkova, E. (2026) Multi-view recipes for making sense of multi-source data originating from fleets of industrial assets. IEEE Access. 

Autonomous vehicles experience difficulties when the environment changes because of fog, night, or glare. FAIR researchers built systems that recognize these shifts and adjust their perception models in real time, improving safety and reliability.

A multi-modal perception framework combining radar, LiDAR and camera inputs with a ResNet-based context estimator improves object detection by 8.8% average precision across varying weather and lighting conditions (KITTI, NuScenes). This makes perception more robust for autonomous driving and industrial robotics.

Robots and autonomous systems often make mistakes not because they lack intelligence, but because they react too slowly. A team at KU Leuven built a new type of hybrid AI chip that integrates perception, reasoning and decision‑making on one compact device. By processing sensing and thinking at the same time, these chips allow robots to respond almost instantly.

Running the full neuro-symbolic pipeline (neural networks for perception combined with symbolic logic for robust reasoning and planning) on the Hemaia chip resulted in up to 115× lower end-to-end latency compared with a standard edge GPU. The chip integrates a Tensor Core for sensory processing, a Hyperdimensional Computing (HDC) engine for robust symbolic manipulation, and a CGRA for structured transforms. The chip supports low-power reasoning and is designed for safety-critical edge applications such as industrial robots and autonomous vehicles.

FAIR researchers developed AI controllers that constantly adjust turbine operations, hydrogen production and market participation in offshore wind farms to maximize value.

Multi-agent reinforcement learning coordinates a wind farm and electrolyser wind-to-hydrogen systems end-to-end. Tests show significant increases in operational revenue relative to conventional rule-based controllers, while smoothing renewable integration and responding to price fluctuations in real time.

Example papers:

• Ally, S., T. Verstraeten, P.-J. Daems, A. Nowé, and J. Helsen (2025). Modular deep learning approach for wind farm power forecasting and wake loss prediction. Wind Energy Science, 10(4):779–812.

• Ally, S., T. Verstraeten, A. Nowé, and J. Helsen (2025). Day-ahead trading and power control for hybrid wind-hydrogen plants with multi-agent reinforcement learning. Applied Energy, 401:126588.

Traditional sleep studies require overnight hospital stays. FAIR researchers instead used simple home recordings using a microphone and oxygen sensor to assess sleep apnea reliably. This has the potential to make large‑scale screening much more comfortable and routine.

RSN-Count represents a paradigm shift in automated machine learning-based sleep apnea assessment by prioritizing event counting over precise localization, the model estimates the Apnea-Hypopnea Index with a 25–30% reduction in error. The model uses spiking neural networks for efficient sequence processing and is viable for home-based clinical screening.