Research

My core field: how local rules and interactions on a network give rise to collective behavior — consensus, dissent, avalanches, and the structure that steers them.

Most of my work starts from one question: how do local and simple rules — where each node or agent sees only its own neighborhood — produce behavior we recognize as collective? I pursue this along two lines.

Dynamics on networks.

I am investigating how actors reach a consensus, or, to put it more accurately, how actors arrive at a state in which they believe they have reached a consensus. My “polycontextural networks” consist of actors with purely subjective perceptions and no shared value system. Through various mechanisms, the network settles into stable, internally consistent states that do not represent consensus at all (“stable misunderstandings”). A single reaction parameter controls a phase transition between a phase of constant opinion formation and a frozen phase, with self-similar, convention-oriented clusters at the critical point: Collective patterns and stable misunderstandings; Nucleation transitions in polycontextural networks

Which connections matter.

Given a process running on a graph, which edges carry the dynamics — insulating regions, or promoting spread? I analyzed this for a model (Graph Coloring Dynamics), where the Agents aim to differentiate from each other. Structural insulators and promotors in networks. We also built a method to score how much each edge matters for a given dynamics, and found that excitable spreading and pattern formation rely on different parts of the same graph. Applied to the macaque cortex, the important edges align with biological function. Categories of important edges in dynamics on graphs.

Current direction (work in progress):

I am extending the polycontextural network framework toward systems that observe and represent one another — a first step toward modeling self-reflexive behavior. This research is closely intertwined with my work on machine learning systems and my theoretical work on polycontextural logic (see area 02).

What does it mean for a system — or a model — to take a perspective and observe itself? Foundational work on polycontextural logic in physics, and a frontier project on self-reflexive systems.

This is where I step back from specific models to ask foundational questions about how perspectives, observation, and representation work — and where my most forward-looking project sits. It connects directly to the network research in area 01.

Polycontextural logic and physics.

Some perspectives are irreducible: they cannot always be collapsed into a single objective, view-from-nowhere description. Building on Gotthard Günther’s polycontextural logic — a framework for reasoning across several mutually incompatible viewpoints at once — I proposed that physical theories are best understood as the transformations that connect distinct contextures, rather than as one unified account handed down from outside. Physics is organized around transformations connecting contextures in a polycontextural world.

Toward self-reflexive systems (work in progress).

A central idea in Günther’s logic is that self-reflection arises from mutual observation — agents become able to take a perspective on themselves by observing one another. My current direction makes this operational: using the polycontextural networks from area 01 as a substrate for systems that mutually observe and represent each other, with the longer-term aim of implementing self-reflexive behavior in interacting artificial and learning systems. This work is early-stage and as yet unpublished; the consensus-network papers in area 01 lay the groundwork.

How randomness shapes dynamical systems — noise that creates and destroys stable states and bifurcations, and the theory used to describe it.

Stochasticity, as it occurs in most real-world systems, can do far more than add a small blur on top of deterministic behavior — it can change that behavior qualitatively. This line of work asks how noise creates, removes, and reshapes the stable states and transitions of a dynamical system. The application domain is often molecular and biological, but the methods — master equations, stochastic dynamics, bifurcation analysis — are general.

Noise that creates, not just blurs.

In a minimal chemical model, I kept the deterministic equations fixed and varied only how molecules are produced in bursts. That change alone can create or destroy a bistable switch: the noise is not blurring the dynamics, it is generating them. A minimal model of burst-noise induced bistability. To analyze such systems systematically, we extended the familiar fixed-points-and-flows picture to noisy dynamics with stochastic phase portraits, which revealed a new type of bifurcation where the most likely state escapes through a gap between the nullclines. Analysis of stochastic bifurcations with phase portraits.

Systems in context.

A dynamical system is rarely isolated. Using a largely master-equation–based analysis, we showed that even simple mono-molecular reactions in the surrounding environment endow an embedded reaction system with a form of memory: its present behavior depends on its past in ways that are invisible when the system is studied on its own. Context in synthetic biology: memory effects of environments with mono-molecular reactions.

From model to engineered circuit.

I applied the same burst-noise modeling approach to design the model of a synthetic gene circuit — a tightly regulated, tunable CRISPR-dCas9 AND gate — in collaboration with experimentalists, who built and measured the system. My contribution was the model; the joint work appeared in Nucleic Acids Research. A tightly regulated and adjustable CRISPR-dCas9 based AND gate in yeast.

Applied ML on real, messy data — where a leaderboard and a hard deadline test an idea fast. Competition wins and collaborative modeling, kept interpretable wherever possible.

Alongside the theoretical work, I apply machine learning to concrete data problems.

Competitions.

I lead teams in international data-science challenges: first place in the PEGS DREAM Challenge 2024 — classifying hypercholesterolemia from combined genomic, exposome, and geospatial data — and back-to-back wins at the Bremen Big Data Challenge 2024 and 2025, including a fraud-detection model.

In practice this is supervised learning on tabular and multimodal data — gradient boosting, random forests, ensemble methods. However, I always put an emphasis on interpretability and feature selection over black-box performance alone. This approach sets my work apart from many others.

Industry projects.

Since 2019 I have carried methods from complex-systems science and ML into industry-funded projects at Constructor University Bremen. Specifically, in 2025 and 2026, I analyzed data for a leading manufacturer of mass spectrometers and provided reports that will lead to improved performance and more efficient maintenance in the long term. In 2026, I also created a dashboard that allows the company to analyze recorded measurement data. It was an exciting opportunity for me to work with tools like FastAPI and SQLModel. The dashboard is now in production.