Cover of the PhD thesis. (Image by L. Natali.)Laura Natali defended her PhD thesis on March 28th, 2025. Congrats!
The defense took place in PJ, Institutionen för fysik, Origovägen 6b, Göteborg, at 10:00.
Title: Neural Networks for Complex Systems: From Epidemic Modeling to Swarm Robotics
Abstract: Deep learning models, inspired by the structure of the brain, were first developed in the last century. These models are trained to recognize patterns in large amounts of data. Recently, deep learning has made a big impact, both in research and in everyday applications, like healthcare, image recognition, and language translation.
However, despite their advancements, these models still fall short of the abilities found in biological brains, which are adaptable, energy-efficient, and have evolved over millions of years. In contrast, artificial models are specialized and struggle to adapt to new information.
To help address this gap, we have developed a robotic experiment that combines the programmability of artificial neural networks with some of the physical constraints seen in biological systems.
Supervisor: Giovanni Volpe Examiner: Bernhard Mehlig Opponent: Hamid Kellay Committee: Maria Guix Noguera, Juliane Simmchen, Michael Felsberg Alternate board member: Paolo Vinai
From left: Anupam Sengupta (opponent), Harshith Bachimanchi, Giovanni Volpe (supervisor). (Photo by A. Argun.)Harshith Bachimanchi defended his PhD thesis on March 26th, 2025. Congrats!
The defense took place in PJ, Institutionen för fysik, Origovägen 6b, Göteborg, at 13:00.
Title: Deep Learning Enhanced Optical Methods for Single-Plankton Studies
Abstract: Among Earth’s earliest life forms, cyanobacteria reshaped the planet by oxygenating the atmosphere during the Great Oxidation Event 2.4 billion years ago. This process, which led to ozone formation and UV protection, paved the way for more complex photosynthetic organisms—phytoplankton, the eukaryotic descendants of cyanobacteria. Today, phytoplankton drive the global carbon cycle, producing 50–80% of Earth’s oxygen and fueling the marine food web. Microzooplankton consume nearly two-thirds of the organic carbon generated, yet despite their ecological significance, tracking biomass flow at the single-cell level remains a major challenge.
This thesis presents novel methodologies that integrate advanced optical techniques, deep learning, and simulated datasets to analyze microplankton dynamics with unprecedented resolution.
A key contribution is a deep-learning-enhanced holographic microscopy approach that quantifies microplankton biomass at the single-cell level while simultaneously capturing their three-dimensional swimming behavior. This method overcomes computational bottlenecks in traditional holography, enabling high-throughput analysis across diverse species and size ranges. Expanding on this, I demonstrate its application in mixed-species experiments to examine feeding interactions between phytoplankton and microzooplankton, capturing biomass transfer and behavioral shifts during predation.
Beyond direct imaging, this thesis leverages synthetic data to advance microscopy-based research. Neural networks trained on simulated microscopy datasets are used to detect, segment, and classify plankton species while reconstructing motion dynamics. To showcase the versatility of this approach, I present its application in a non-biological setting—detecting bubble-propelled artificial micromotors within complex experimental backgrounds. In addition to object detection, these methods also enable motion characterization of microscopic entities. To demonstrate this, I introduce synthetic microscopy videos that model microscopic organisms undergoing various anomalous diffusion behaviors. This framework is then used to develop a method that extracts motion characteristics without explicit trajectory linking, broadening its applications beyond plankton ecology.
Finally, I investigate how zooplankton—key players in the marine food web—respond to ocean wave-induced light patterns using an LED matrix. The results suggest that zooplankton use steady light sources, such as celestial objects, to ascend more rapidly during favorable low-turbulent conditions, offering new insights into their migratory strategies. Collectively, this thesis bridges marine ecology, microscopy, artificial intelligence, and biophysics to provide new tools for exploring the unseen dynamics that shape our planet.
Memory capacity in aging. A Brain reservoir computing architecture with uniform random signals applied to all nodes. (Image from the article.)Computational memory capacity predicts aging and cognitive decline
Mite Mijalkov, Ludvig Storm, Blanca Zufiria-Gerbolés, Dániel Veréb, Zhilei Xu, Anna Canal-Garcia, Jiawei Sun, Yu-Wei Chang, Hang Zhao, Emiliano Gómez-Ruiz, Massimiliano Passaretti, Sara Garcia-Ptacek, Miia Kivipelto, Per Svenningsson, Henrik Zetterberg, Heidi Jacobs, Kathy Lüdge, Daniel Brunner, Bernhard Mehlig, Giovanni Volpe, Joana B. Pereira
Nature Communications 16, 2748 (2025)
doi: 10.1038/s41467-025-57995-0
Memory is a crucial cognitive function that deteriorates with age. However, this ability is normally assessed using cognitive tests instead of the architecture of brain networks. Here, we use reservoir computing, a recurrent neural network computing paradigm, to assess the linear memory capacities of neural-network reservoirs extracted from brain anatomical connectivity data in a lifespan cohort of 636 individuals. The computational memory capacity emerges as a robust marker of aging, being associated with resting-state functional activity, white matter integrity, locus coeruleus signal intensity, and cognitive performance. We replicate our findings in an independent cohort of 154 young and 72 old individuals. By linking the computational memory capacity of the brain network with cognition, brain function and integrity, our findings open new pathways to employ reservoir computing to investigate aging and age-related disorders.
GAUDI leverages a hierarchical graph-convolutional variational autoencoder architecture, where an encoder progressively compresses the graph into a low-dimensional latent space, and a decoder reconstructs the graph from the latent embedding. (Image by M. Granfors and J. Pineda.)Global graph features unveiled by unsupervised geometric deep learning Mirja Granfors, Jesús Pineda, Blanca Zufiria Gerbolés, Joana B. Pereira, Carlo Manzo, Giovanni Volpe
arXiv: 2503.05560
Graphs provide a powerful framework for modeling complex systems, but their structural variability makes analysis and classification challenging. To address this, we introduce GAUDI (Graph Autoencoder Uncovering Descriptive Information), a novel unsupervised geometric deep learning framework that captures both local details and global structure. GAUDI employs an innovative hourglass architecture with hierarchical pooling and upsampling layers, linked through skip connections to preserve essential connectivity information throughout the encoding-decoding process. By mapping different realizations of a system – generated from the same underlying parameters – into a continuous, structured latent space, GAUDI disentangles invariant process-level features from stochastic noise. We demonstrate its power across multiple applications, including modeling small-world networks, characterizing protein assemblies from super-resolution microscopy, analyzing collective motion in the Vicsek model, and capturing age-related changes in brain connectivity. This approach not only improves the analysis of complex graphs but also provides new insights into emergent phenomena across diverse scientific domains.
Recent advances in nonequilibrium physics allow extracting thermodynamic quantities, such as entropy production, directly from dynamical information in microscopic movies. (Image by S. Manikandan.)Localizing entropy production in non-equilibrium processes
Sreekanth Manikandan Date: 5 Mar 2025 Time: 14:45 Place: Nordita
Part of the 14th Nordic Workshop on Statistical Physics
Quantifying the spatiotemporal forces, affinities, and dissipative costs of cellular-scale non-equilibrium processes from experimental data and localizing it in space and time remain a significant open challenge. Here, I explore how principles from stochastic thermodynamics, combined with machine learning techniques, offer a promising approach to addressing this issue. I will present preliminary results from experiments on fluctuating cell membranes and simulations of non-equilibrium systems in stationary and time-dependently driven states. These studies reveal potential strategies for localizing entropy production in experimental biophysical contexts while also highlighting key challenges and limitations that must be addressed.
Propagation of scattered light through a scattering microscope, illustrating typical nanoparticles studied. (Image by B. García Rodriguez.)Optical Label-Free Microscopy Characterization of Dielectric Nanoparticles
Berenice Garcia Rodriguez, Erik Olsén, Fredrik Skärberg, Giovanni Volpe, Fredrik Höök, Daniel Sundås Midtvedt
Nanoscale, 17, 8336-8362 (2025)
arXiv: 2409.11810
doi: 10.1039/D4NR03860F
In order to relate nanoparticle properties to function, fast and detailed particle characterization, is needed. The ability to characterize nanoparticle samples using optical microscopy techniques has drastically improved over the past few decades; consequently, there are now numerous microscopy methods available for detailed characterization of particles with nanometric size. However, there is currently no “one size fits all” solution to the problem of nanoparticle characterization. Instead, since the available techniques have different detection limits and deliver related but different quantitative information, the measurement and analysis approaches need to be selected and adapted for the sample at hand. In this tutorial, we review the optical theory of single particle scattering and how it relates to the differences and similarities in the quantitative particle information obtained from commonly used microscopy techniques, with an emphasis on nanometric (submicron) sized dielectric particles. Particular emphasis is placed on how the optical signal relates to mass, size, structure, and material properties of the detected particles and to its combination with diffusivity-based particle sizing. We also discuss emerging opportunities in the wake of new technology development, with the ambition to guide the choice of measurement strategy based on various challenges related to different types of nanoparticle samples and associated analytical demands.
MIRO employs a recurrent graph neural network to refine SMLM point clouds by compressing clusters around their center, enhancing inter-cluster distinction and background separation for efficient clustering. (Image by J. Pineda.)Relational Inductive Biases as a Key to Smarter Deep Learning Microscopy Jesús Pineda Learning on graphs and geometry meetup at Uppsala University Date: 12 February 2025 Time: 11:15 Place: Lecture hall 4101, Ångströmlaboratoriet, Uppsala, Sweden
Geometric deep learning has revolutionized fields like social network analysis, molecular chemistry, and neuroscience, but its application to microscopy data analysis remains a significant challenge. The hurdles stem not only from the scarcity of high-quality data but also from the intrinsic complexity and variability of microscopy datasets. This presentation introduces two groundbreaking geometric deep-learning frameworks designed to overcome these barriers, advancing the integration of graph neural networks (GNNs) into microscopy and unlocking their full potential. First, we present MAGIK, a cutting-edge framework for analyzing biological system dynamics through time-lapse microscopy. Leveraging a graph neural network augmented with attention-based mechanisms, MAGIK processes object features using geometric priors. This enables it to perform a range of tasks, from linking coordinates into trajectories to uncovering local and global dynamic properties with unprecedented precision. Remarkably, MAGIK excels under minimal data conditions, maintaining exceptional performance and robust generalization across diverse scenarios. Next, we introduce MIRO, a novel algorithm powered by recurrent graph neural networks. MIRO pre-processes Single Molecule Localization (SML) datasets to enhance the efficiency of conventional clustering methods. Its ability to handle clusters of varying shapes and scales enables more accurate and consistent analyses across complex datasets. Furthermore, MIRO’s single- and few-shot learning capabilities allow it to generalize effortlessly across scenarios, making it an efficient, scalable, and versatile tool for microscopy data analysis. Together, MAGIK and MIRO address critical limitations in microscopy data analysis, offering innovative solutions for multi-scale data analysis and advancing the boundaries of what is currently achievable with geometric deep learning in the field.
GAUDI leverages a hierarchical graph-convolutional variational autoencoder architecture, where an encoder progressively compresses the graph into a low-dimensional latent space, and a decoder reconstructs the graph from the latent embedding. (Image by M. Granfors and J. Pineda.)Global graph features unveiled by unsupervised geometric deep learning
Mirja Granfors, Jesús Pineda, Blanca Zufiria Gerbolés, Daniel Vereb, Joana B. Pereira, Carlo Manzo, and Giovanni Volpe Learning on graphs and geometry meetup at Uppsala University Date: 11 February 2025 Place: Uppsala university
Graphs are used to model complex relationships, such as interactions between particles or connections between brain regions. The structural complexity and variability of graphs pose challenges to their efficient analysis and classification. Here, we propose GAUDI (Graph Autoencoder Uncovering Descriptive Information), a graph autoencoder that addresses these challenges. GAUDI’s encoder progressively reduces the size of the graph using multi-step hierarchical pooling, while its decoder incrementally increases the graph size until the original dimensions are restored, focusing on the node and edge features while preserving the graph structure through skip-connections. By training GAUDI to minimize the difference between the node and edge features of the input graph and those of the output graph, it is compelled to capture the most critical parameters describing these features in the latent space, thereby enabling the extraction of essential parameters characterizing the graphs. We demonstrate the performance of GAUDI across diverse graph data originating from complex systems, including the estimation of the parameters of Watts-Strogatz graphs, the classification of protein assembly structures from single-molecule localization microscopy data, the analysis of collective behaviors, and correlations between brain connections and age. This approach offers a robust framework for efficiently analyzing and interpreting complex graph data, facilitating the extraction of meaningful patterns and insights across a wide range of applications.
(Photo by A. Ciarlo.)Ade Satria Saloka Santosa joins the Physics Department of the University of Gothenburg as a visiting PhD student from Uppsala University on 1 February 2025.
Ade holds a Master of Science degree in Industrial Chemistry from Pukyong National University, South Korea, and has research experience at the Korea Institute of Materials Science (KIMS).
During his PhD, he will focus on nanofabrication and e-paper technology.
MIRO employs a recurrent graph neural network to refine SMLM point clouds by compressing clusters around their center, enhancing inter-cluster distinction and background separation for efficient clustering. (Image by J. Pineda.)Spatial clustering of molecular localizations with graph neural networks
Jesús Pineda Date: 27 January 2025 Time: 10:00 Place: SciLifeLab Campus Solna, Sweden
Single-molecule localization microscopy (SMLM) generates point clouds corresponding to fluorophore localizations. Spatial cluster identification and analysis of these point clouds are crucial for extracting insights about molecular organization. However, this task becomes challenging in the presence of localization noise, high point density, or complex biological structures. Here, we introduce MIRO (Multimodal Integration through Relational Optimization), an algorithm that uses recurrent graph neural networks to transform the point clouds in order to improve clustering efficiency when applying conventional clustering techniques. We show that MIRO supports simultaneous processing of clusters of different shapes and at multiple scales, demonstrating improved performance across varied datasets. Our comprehensive evaluation demonstrates MIRO’s transformative potential for single-molecule localization applications, showcasing its capability to revolutionize cluster analysis and provide accurate, reliable details of molecular architecture. In addition, MIRO’s robust clustering capabilities hold promise for applications in various fields such as neuroscience, for the analysis of neural connectivity patterns, and environmental science, for studying spatial distributions of ecological data.
