Research can sound complex, but its purpose is simple
At MetaInfrastructure, we explore how the places and systems we all rely on — such as cities, bridges, transport, and energy networks — can be safer, fairer, and better prepared for the future.
We study how infrastructure is affected by climate change, natural hazards, and rapid urban growth, and we develop practical tools, ideas, and technologies to make these systems more resilient, sustainable, and people-centric. Our work combines engineering with digital technologies such as artificial intelligence, data modelling and smart sensors to help decision-makers tackle real challenges and improve everyday life for communities around the world.
Meet our Research:
The Crowd knows best: ensemble-based learning for complex civil engineering applications
by Ivan Izonin, Associate Professor & Research Fellow
Ensemble learning is an approach in which multiple expert (weak learner) models are combined to obtain a more accurate and reliable prediction than any single model can provide. It operates by aggregating diverse decisions that partially compensate for one another’s individual errors. A key requirement is sufficient diversity and at least partial independence among the errors of the constituent models.
The conceptual foundation of ensemble learning traces back to the “wisdom of the crowd” principle (Galton, 1907), which showed that the averaged judgment of a large group of independent “experts” can surpass the accuracy of individual participants. In the context of machine learning, this means that combining several models, each offering a different interpretation of the data structure, leads to a more robust and dependable prediction than relying on any one model. A more formal justification comes from Condorcet’s jury theorem (Condorcet, 1785; later adapted for ML), which states that if each expert’s probability of being correct is slightly above random, then collective voting substantially increases overall accuracy as the number of experts grows. In machine learning, this explains why ensembles of many weak learners can outperform a single strong learner—provided that their errors are sufficiently diverse and only partially correlated.
Additional theoretical grounding is offered by the bias–variance decomposition (Geman et al., 1992), which demonstrates that averaging predictions across multiple models significantly reduces variance, i.e., fluctuations caused by sampling noise. At the same time, the increase in bias is minimal or negligible. As a result, ensembles produce more reliable, generalizable estimates, explaining their effectiveness across a broad range of tasks – from classical regression to complex high-dimensional classification problems.
Making Cities and Infrastructure Safer from Wildfires
by Stavros Sakellariou, Research Fellow
Wildfires are becoming more frequent and intense due to climate change, and they no longer affect only forests. Increasingly, fires threaten cities, homes, and critical infrastructure such as roads, power networks, and water systems—especially in areas where urban development meets vegetation. My research focuses on understanding how wildfires interact with cities and infrastructure, and how we can better prepare for and reduce their impacts.
This problem matters because wildfire impacts go far beyond burned land. They can disrupt daily life, endanger public health through smoke, damage essential services, and lead to costly evacuations and long recovery times. As more people live close to fire-prone landscapes, communities face growing risks to safety, economic stability, and quality of life.
My research helps by combining spatial data, fire simulations, and infrastructure analysis to identify where risks are highest and which assets are most vulnerable. By translating complex wildfire behaviour into maps and indicators, the work supports planners, emergency services, and decision-makers in choosing smarter prevention measures, safer urban designs, and more effective response strategies—before disasters happen.