When
Where
Speaker: Brenda Karime Alvarez Ortiz
Title: Learning the Hidden Dynamics of Biosphere 2: From Data Geometry to Predictive Control
Abstract: Biosphere 2 is a highly complex, closed ecological system where internal climate dynamics arise from the interaction of solar forcing, thermal inertia, humidity, and mechanical control. Traditional modeling approaches based on explicit physical equations or linear time series methods often fail to capture the system’s behavior during critical transitions, such as periods of extreme thermal load.
In this talk, I present a data-driven approach to uncovering the underlying dynamical structure of the system directly from sensor data. Using synchronized multivariate time series, sliding-window feature extraction, and dimensionality reduction via Principal Component Analysis, we observe that the system evolves on a low-dimensional manifold despite the presence of many interacting variables. Clustering methods are then used to identify distinct behavioral regimes, which correspond to different thermodynamic states of the biosphere.
These results provide a bridge between data-driven geometry and physics-based modeling, and connect naturally to Markov-switching state-space frameworks for predictive control. Ultimately, this work contributes toward the development of a behavioral digital twin capable of anticipating regime transitions and enabling proactive climate control strategies in Biosphere 2.
Speaker: Tim Zou
Title: Network Structure and Neuronal Synchrony in Spiking Neural Networks
Abstract: Understanding how different neuronal network architectures influence network dynamics is a fundamental challenge in neuroscience. In this study, we investigate how local connectivity in ring-structured neuronal networks shapes the emergence of synchronous firing events. The ring model, a widely used mathematical abstraction of cortical networks, consists of excitatory and inhibitory neurons arranged on a circle, with each neuron projecting to its nearest $NC$ neighbors. This localized connectivity can lead to complex synchronization patterns, but the precise relationship between network structure and collective neuronal behavior remains poorly understood.
To address this, we simulate the dynamics of a 600-neuron integrate-and-fire ring model with varying local connectivity parameters. Using coarse-grained analysis, we identify distinct firing regimes, including homogeneous, intermediate, and synchronized phases, each characterized by different degrees of neuronal correlation and cluster formation. Our results include a mean-field estimation of the average firing rate within each regime, as well as a theoretical framework for predicting the critical points where the network transitions between these phases. These findings provide new insights into how local network architecture influences global neuronal dynamics, offering a quantitative basis for understanding the emergence of synchrony in biological networks. Additionally, this work highlights the critical role of fixed, localized connections in shaping the dynamics of spiking neural networks, potentially informing the design of more biologically realistic models.