Projects & Tutors
1) Inferring empirical belief networks to ground agent-based models of polarisation
The first aim of the project is the inference of a belief space from political survey data, where participants indicate their agreement level to a set of policy issues. This can be linked to the broader problem of network inference from data (or the inverse Ising problem) and the psychology sub-field of belief networks. Once a data-driven belief space has been inferred, the second aim is to use this belief space to calibrate and run agent-based simulations of a multi-dimensional opinion dynamics model, and study what kind of outcomes (consensus or polarization) emerge from this empirical belief structure.
As opinion dynamics model, we will use the multi-dimensional opinion dynamics model by Mueller & Tan. Their model is simple yet elegant, as they showed that assuming conceptual coherence and restrictions between related beliefs is enough to create a model that is able to lead to fragmentation and polarisation to the extremes without a need of repulsive forces or bounded confidence. Their model is also strongly backed by cognitive science theory and knowledge from experimental results, but no calibration or validation with data exists to this date. By using a political attitude dataset and calibrating the model on this dataset, we will bridge the gap between the theoretical model and empirical cases.
This task requires different sets of skills, although these can be distributed among the team members. As a team, they should have: some basic data wrangling (categorical data), basic knowledge of probability inference, computational knowledge for working with data with medium size, network science for understanding and analysing belief networks, experience at building Agent-Based Models, and, ideally, domain knowledge (political) to interpret the outcomes. Depending on the team expertise and interests, the focus can be on only some parts of the proposed project.
Guillermo Romero Moreno
Guillermo Romero Moreno is a postdoctoral researcher at the Centrum Wiskunde & Informatica in Amsterdam and within the AI, Media and Democracy Lab, where he is studying the spread dynamics of fake news in social media platforms with the goal of testing the effectiveness of alternative interventions. Before coming to Amsterdam, Guillermo conducted postdoctoral research at the University of Edinburgh on Bayesian inference of multimorbidity networks and completed his PhD in opinion dynamics at the University of Southampton. His research approach is greatly interdisciplinary, between Computational Social Sciences, complex systems, and computational methods.
Adarsh Prabhakaran
2) Linking the “inner” and “outer” self to brain networks
Recent research has demonstrated that the organization of brain networks varies across different experiences of the social world. For example, there is increased connectivity within the Default Mode Network among people with greater social support. Similarly, researchers have found that brain network connectivity predicts how sociable a person will be. How do the neural underpinnings of these “inner” and “outer” social selves compare? For this Complexity 72h project, we aim to leverage data from the Human Connectome Project to address this question. The Human Connectome Project contains measures of social support, social behaviors, social cognition, mental health, and more for 1200 people. Additionally, it includes resting state fMRI data to construct functional brain networks. With these data, we will address how the network neural associations of “inner” social self phenomena (e.g. sociability, extroversion) and “outer” social self phenomena (e.g. social support, experiences of friendliness) overlap and diverge across different brain systems. More than this, we will cluster individuals based on these “inner” and “outer” social self measures to identify different social profiles and determine the extent to which associations with connectivity are dissociable across these different social profiles. Finally, given the strong association between social support and mental health, we will assess the three-way associations between mental health, brain network connectivity, and social profiles. This line of questioning has interesting philosophical consequences for understanding how the social world shapes the self and brain network organization, as well as how we characterize “inner” versus “outer” social and personal experiences.
Haily Merritt
Haily is a multidisciplinary scientist who studies how social relationships shape brain networks, mental health, and behavior using network science and complex-systems methods. She is a Postdoctoral Research Assistant with Dr. Sofia Teixeira at the Network Science Institute (Northeastern University London) and earned a dual PhD from Indiana University in Informatics (Complex Networks and Systems) and Cognitive Science. She also serves as DEI Coordinator for Women in Network Science (WiNS) and helps organize the WiNS Seminar Series. Outside academia, she enjoys reading, painting, and gardening.
Andreia Sofia Teixeira
Andreia is an Associate Professor (Teaching and Research) at the Network Science Institute, Northeastern University London, where she leads the BRAN Lab. She is also a Visiting Academic at Kent and Medway Medical School and an Academic Honorary Researcher at the UK Health Security Agency, and she serves on the Executive Committee of the Network Science Society as well as Vice President of Women in Network Science.
Her research sits at the intersection of network science and machine learning, developing measures, computational models, and simulation frameworks to study complex systems across domains including graph theory, computational cognitive/social science, network neuroscience, and computational medicine/epidemiology. She is particularly interested in collective behavior, neurodegenerative and neuropsychiatric disorders, and how affective relationships shape (and are shaped by) our environments.
Carlos Gershenson
Prof. Gershenson is a tenured full professor at SUNY Binghamton. He is also affiliated with the the Centro de Ciencias de la Complejidad and the Instituto de Investigaciones en Matemáticas Aplicadas y en Sistemas of the Universidad Nacional Autónoma de México (UNAM). He has a broad range of academic interests, including complexity, artificial intelligence, artificial life, healthcare, self-organizing systems, information, urbanism, evolution, cognition, and philosophy of science.
He is Vicepresident Secretary of the Complex Systems Society (2021-) Editor-in-Chief of Complexity Digest (2007-), and member of the Board of Advisors for Scientific American (2018-).
He was previously a Research Professor at UNAM (2008-2023), and visiting professor/researcher/scholar at the Santa Fe Institute (2022-2023), MIT (2015-2016), Northeastern (2015-2016), and ITMO University (2015-2019).
3) Predicting Amazonian biodiversity from canopy height variation
We propose to develop a predictive model of Amazonian biodiversity using lidar-derived canopy height data from the CTrees Amazon Canopy Height dataset. Structural heterogeneity in forest canopies—quantified as height variation—is a well-established indirect proxy for biodiversity: forests with greater vertical complexity typically support higher species richness among trees, plants, birds, amphibians, and arthropods.
Our approach will quantify canopy height variability and gap structure metrics from high-resolution CHMs to predict multiple aspects of biodiversity, including plant and animal species richness and abundance. We will leverage machine learning methods to link structural heterogeneity with field-based biodiversity data, aiming to identify predictive patterns and thresholds where canopy complexity most strongly correlates with species diversity.
Beyond biodiversity, our model will provide insights into key ecological processes. Canopy gaps influence forest dynamics by driving recruitment, mortality, and regeneration pathways, while structural heterogeneity affects microclimate buffering—important for understory temperature regulation, habitat suitability, and climate refugia. Additionally, canopy structure shapes ecosystem processes such as carbon cycling, where decomposition rates differ markedly between gaps and closed canopy areas.
By integrating structural metrics with ecological outcomes, this project will generate spatially explicit predictions of biodiversity and forest function across Amazonia. The results will advance understanding of the “Height Variation Hypothesis” in tropical forests, support conservation planning, and inform climate-resilient management strategies by linking forest structure, biodiversity, and ecosystem functioning.
Given the interdisciplinary nature of this project, ideal team members would collectively bring expertise across different disciplines.
Fakhteh Ghanbarnejad
Fakhteh is currently a Professor of Computer Science at SRH University of Applied Sciences. She received my B.Sc. and M.Sc. from the Department of Physics at Sharif University of Technology (SUT), one of the top universities in the Middle East, and hold a Ph.D. in Bioinformatics from Leipzig University in Germany. After completing a postdoc at the Max Planck Institute for the Physics of Complex Systems (MPIPKS) in Dresden, Germany, she worked as a Principal Investigator (PI) at the Technical Universities of Berlin (TUB) and Dresden (TUD). Fakhteh has also held positions as an Assistant Professor in Physics at SUT, a Research Staff Associate in the Quantitative Life Sciences section at the Abdus Salam International Centre for Theoretical Physics (ICTP) in Trieste, Italy, and as a Senior Researcher at the Robert Koch Institute (Public Health), the Helmholtz Centre for Infection Research (Public Health), and the Potsdam Institute for Climate Impact Research (PIK) in Germany.
Jan Nagler
Jan Nagler studied physics in Kiel, Germany, before he went for his PhD (on dynamical systems, in particular problems in chaotic astrophysics) to Bremen, Germany. His scientific postdoctoral positions include Boston University, USA (research on networks and econophysics), the Max-Planck-Institute for Dynamics and Self-Organization, Goettingen, Germany, and ETH Zurich & ETH Risk Center, Switzerland.
From 2016 to 2022 he was the Vice Chair of the Physics of Socio-economic Systems Division of the German Physical Society.
His research includes the understanding and control of networked stochastic systems, with applications at the interface between physics, biology, and socio-economic systems, in particular game theory and phase transitions, ergodicity breaking and risk and survival in uncertain environments.
Jan Nagler has been the Director of the Center for Human and Machine Intelligence since 2023.
4) Regional evolution of greenhouse-gas growth rates from multi-satellite observations
Greenhouse-gas (GHG) concentrations have been rising steadily since the industrial revolution, and many countries have pledged to reduce emissions under the Paris Agreement. A key scientific and societal challenge is determining whether such reductions are actually detectable in the atmosphere. This is difficult because natural variability—such as year-to-year changes in biospheric CO₂ uptake—can obscure or mimic emission-driven trends. Disentangling human and natural influences therefore requires analytical approaches capable of handling complex, high-dimensional data.
A wide range of satellite missions now provide global measurements of atmospheric composition, including total-column CO₂ and CH₄, as well as co-emitted pollutants like NO₂ and CO that trace fossil-fuel combustion. Additional products such as solar-induced chlorophyll fluorescence (SIF) offer insight into biospheric activity. These observations are complemented by bottom-up emission inventories and model-based reanalysis datasets, together forming a diverse set of information on the carbon cycle.
This project will integrate these datasets within a machine-learning and complex-systems framework to explore how GHG trends evolve across different regions. Students will investigate questions such as: Where do CO₂ and CH₄ growth rates show unexpected changes? Which patterns are driven by emissions, and which arise from natural variability? Are there emerging anomalies or regional behaviours that standard analyses overlook? This interdisciplinary project is suitable for students from atmospheric science, physics, computer science, or related fields who are interested in working with real-world environmental data and modern analytical tools.
Fakhteh Ghanbarnejad
Fakhteh is currently a Professor of Computer Science at SRH University of Applied Sciences. She received my B.Sc. and M.Sc. from the Department of Physics at Sharif University of Technology (SUT), one of the top universities in the Middle East, and hold a Ph.D. in Bioinformatics from Leipzig University in Germany. After completing a postdoc at the Max Planck Institute for the Physics of Complex Systems (MPIPKS) in Dresden, Germany, she worked as a Principal Investigator (PI) at the Technical Universities of Berlin (TUB) and Dresden (TUD). Fakhteh has also held positions as an Assistant Professor in Physics at SUT, a Research Staff Associate in the Quantitative Life Sciences section at the Abdus Salam International Centre for Theoretical Physics (ICTP) in Trieste, Italy, and as a Senior Researcher at the Robert Koch Institute (Public Health), the Helmholtz Centre for Infection Research (Public Health), and the Potsdam Institute for Climate Impact Research (PIK) in Germany.
Nasrin Mostafavi Pak
5) The Good, The Bad, and The Liar: Emergent deception in a LLM-enhanced ABM of a sustainability game
The project proposes an extension of an existing, validated LLM-enhanced Agent-Based Model of a sustainability game that could be played by an LLM agent by introducing a communication layer. In the current framework, agents manage the tension between industrial growth and ecological collapse in silence, perceiving only neighbors’ block configurations. I would love to investigate how system dynamics shift when agents exchange messages regarding their intended actions prior to execution. The framework already exists, so the focus will be on collaborative scientific reasoning, experimental design, and result interpretation, aiming to offer participants an high-level research environment which I believe could be a valuable experience.
Specifically, the project will address one or more of the following research questions.
1. Deception as an emergent property: do agents spontaneously misrepresent their intentions (such as promising not to attack while attacking), or is explicit prompting required to trigger manipulative behavior?
2. Permissions effect: how does the global equilibrium change when agents are explicitly informed that lying is a valid strategy?
3. Reputation effect: how does the introduction of memory, which means allowing agents to track discrepancies between past words and actions, alter the overall system behavior?
We already possess the fully functional computational setting used in previous studies and the access to a GPU farm if experiments with local model are necessary, which is ready to be shared and can be modified by participants. Given that the core environment and decision-making architecture are already established, running these experimental scenarios is highly feasible within the 72-hour timeframe, with the goal of producing a preprint.
Francesco Bertolotti
6) Turnovers, synergy, and home advantage: Statistical signatures of success in the NBA
Basketball offers a rich landscape for studying collective dynamics, strategic adaptation, and individual performance under strong competitive constraints. In this project, we propose a data-driven analysis of NBA games to explore complex patterns emerging in the sport. Building on our prior work in sports analytics, we will combine fine-grained match data with statistical and network-based methods to quantify team and player dynamics across multiple temporal scales. Using play-by-play data from all NBA matches over the last four seasons (2021-2025), we will analyze the point-by-point evolution of games to characterize momentum shifts, turnover effects, and quarter-specific strategic responses. At the team level, we will examine performance trajectories across the season, with particular attention to the statistical relevance of home advantage and contextual factors shaping success. At the individual level, we will quantify player performance complementing established basketball metrics with measures able to capture consistency and adaptability. A central focus of the project is the investigation of complementarity and synergy among teammates. By comparing observed lineups and player interactions against appropriate null models, we will test whether certain combinations of players generate statistically significant performance gains beyond individual contributions. We will employ statistical testing and model-based baselines to assess robustness and significance. Depending on participant interest and progress, we might extend empirical work by exploring simple agent-based models of basketball possessions to link observed statistical regularities with plausible generative mechanisms. The project is designed to be accessible to participants with backgrounds in data-driven physics, data science, statistics, or computer science.
Onkar Sadekar
Federico Battiston
7) Universal patterns in egocentric communication networks
Social networks play a central role in shaping human communication, cultural exchange, and community formation. Research has long examined patterns of social interactions and their evolution, with a key focus on the diversity of tie strengths—an indicator of emotional closeness often measured through communication activity. Egocentric networks, comprising an individual (ego) and their contacts (alters), typically exhibit a core of strong ties surrounded by weaker ones, reflecting cognitive and temporal limits on relationship maintenance. Despite this general structure, individuals differ: some concentrate attention on a few close friends, while others distribute it more evenly. These differences remain stable over time, even with high turnover in personal networks, suggesting underlying personality traits or universal mechanisms.
Our analysis of tens of large-scale datasets spanning multiple communication channels (calls, texts, emails, online messages) reveals systematic heterogeneity in tie strength distributions. Most egocentric networks are highly uneven, while a minority show homogeneity. A minimal dynamical model reproduces these patterns, attributing heterogeneity to cumulative advantage (preferential attachment) and homogeneity to random alter choice. A single parameter—preferentiality—captures this balance and remains stable for individuals over time, even as their contacts change. Population-level distributions of preferentiality are strikingly similar across channels.
The next research step is to test whether these patterns extend to offline interactions. We have curated datasets of physical proximity in diverse settings (schools, workplaces, hospitals, events) and collaborative online spaces. Key questions include: Are heterogeneity-homogeneity balances consistent across offline and online contexts? Are individual preferentialities robust across channels? Can our model, which explains up to 88% of egocentric networks per channel, generalize further? Addressing these questions requires analyzing up to 50+ temporal network datasets using Python-based tools for data processing, simulation, and statistical inference. Ideal participants will have expertise in data analysis, network science, and computational modeling.
Gerardo Iñiguez
Dr. Gerardo Iñiguez is associate professor in complex systems and director of TaCoLAB in Tampere University, Finland, and visiting professor in C3-UNAM, Mexico. Gerardo has made fundamental contributions to the understanding of dynamics on networks, namely the interplay between information transfer and social network evolution. His work mixes data analysis with agent-based modelling and covers topics like opinion formation, social contagion, ranking dynamics, and online polarisation. He has led projects in algorithmic bias and tax evasion, the latter funded and implemented by the Mexican government. Gerardo is editor at Interface, ACS, and Frontiers, and founding member of NoSoCSS and ToCSS.
Giulia de Meijere
Giulia de Meijere is a postdoctoral researcher at the TaCoLAB, Tampere University, in Finland. She obtained her PhD in Applied Mathematics from the Gran Sasso Science Institute, in Italy. Giulia studies the propagation of infectious diseases in the presence of partial and evolving adoption of protective behaviours in networked populations. Her current work investigates the structure and dynamics of informational and social (both online and offline) temporal networks, combining data analysis and mathematical modelling.
8) Will everyone become stupid or super-intelligent? The impact of AI on education
Artificial intelligence is rapidly transforming many areas of human life, including education. From early childhood to university, students around the world are increasingly being encouraged to use AI-powered tools due to the promise they offer of making education faster and more personalised. However, the easiest and most accessible way for students to use AI is often to simply ask a chatbot to solve problems. As a result, understanding and anticipating the future impact of these tools has become increasingly urgent.
Empirical studies have looked at how AI affects students’ learning, but the long-term evolution of skills is still largely unknown. Investigating this with real-world data would require extensive, multi-school, longitudinal studies. Simulations, however, allow us to explore plausible future scenarios and the dynamics of skills development.
In this project, we propose a simulation-based study of how students’ educational levels evolve when chatbots are used to support homework and assignments in mathematics and physics. We will model the dynamics of key scientific competencies—such as problem-solving and algebraic skills—within a high-school classroom environment, considering students’ initial abilities, their relationship with AI tools, and peer interactions, ranging from pairwise to higher-order group effects. The model will be grounded in official educational statistics and existing psychological and educational research. Moreover, we will provide a skill grid dataset about the performances of different AI models on selected tasks.
Marta Baratto
Michele Re Fiorentin
Michele Re Fiorentin graduated in physics at the University of Torino (Italy) and got his Ph.D. in theoretical physics from the University of Southampton, UK. He previously worked as a researcher at the Italian Institute of Technology and has been Assistant Professor at the Politecnico di Torino since 2022. His research primarily focuses on first-principles simulations of materials for environmentally friendly applications, on the investigation of the optoelectronic properties of low-dimensional materials. and on the application of machine learning to physical-chemistry problems. He also collaborates on research related to complex systems and epidemic spreading.