Soft Matter Lab

Presentation by A. Lech at SPIE-ETAI, San Diego, 24 August 2026

DeepTrack2: Microscopy Simulations for Deep Learning
Alex Lech, Mirja Granfors, Jiacheng Huang, Benjamin Midtvedt, Jesús Pineda, Harshith Bachimanchi, Carlo Manzo, Giovanni Volpe
Date: 24 August 2026
Time: 11:45 AM – 12:00 PM PDT
Place: Conv. Ctr. Room 2

DeepTrack2 is a flexible and scalable Python library designed to generate physics-based synthetic microscopy datasets for training deep learning models. It supports a wide range of imaging modalities, including brightfield, fluorescence, darkfield, and holography, enabling the creation of synthetic samples that accurately replicate real experimental conditions. Its modular architecture empowers users to customize optical systems, incorporate optical aberrations and noise, simulate diverse objects across various imaging scenarios, and apply image augmentations. Simulations can be executed using a PyTorch backend, enabling GPU acceleration and backpropagation through the optical system, and allowing seamless integration with deep learning workflows. DeepTrack2 is accompanied by a dedicated GitHub page, providing extensive documentation, examples, and an active community for support and collaboration: https://github.com/DeepTrackAI/DeepTrack2.

Invited Presentation by P. Dutta at SPIE-ETAI, San Diego, 24 August 2026

A coarse-grained molecular dynamics framework used to simulate plasmid DNA analyzed via atomic force microscopy (AFM). The resulting images are used to train a U-Net for DNA chain and crossing segmentation and classification. *(Image by P. Dutta.)*

ASAP (AFM simulation and analysis pipeline): a unified deep learning framework for simulation, quality control, and segmentation in atomic force microscopy (Invited Paper)
Prakhar Dutta, Jiacheng Huang, Nazli Demirpehlivan, Thomas Catley, Sylvia Whittle, Carlo Manzo, Rahul Nagshi, Rachel Owen, Giovanni Volpe
Date: 24 August 2026
Time: 10:30 AM – 11:00 AM PDT
Place: Conv. Ctr. Room 2

Atomic force microscopy (AFM) resolves biological structure and mechanics at high resolution, but produces heterogeneous datasets that are noisy and time-consuming to analyze. Although deep learning could automate quality control, segmentation and feature extraction, adoption is limited by scarce ground-truth training data and technical barriers for experimentalists. Here we present ASAP, an open-source tutorial and pipeline implemented in DeepTrack to provide a foundation for AI-enabled AFM. We demonstrate the framework with three examples: (i) an unsupervised variational autoencoder for force-curve quality control; (ii) a dual pathway simulation for DNA, offering both molecular dynamics and rapid, non-MD geometries to generate ground truth for segmentation training; and (iii) a versatile simulation framework that generates synthetic force curves for user-defined surfaces by applying selectable physical models. By consolidating simulation and learning into a modular ecosystem, this work enables users to utilize our pipeline to optimize AFM workflows for efficient data acquisition and processing.

Invited Presentation by G. Volpe at SPIE-Ultrafast Nonlinear Imaging and Spectroscopy XIV, 24 August 2026

DeepTrack 2 Logo. *(Image from DeepTrack 2 Project)*

Artificial Intelligence for Microscopy: From Pixels to Physical Insight
Giovanni Volpe
(Invited Paper)
Date: 24 August 2026
Time: 9:45 AM – 10:10 AM PDT
Place: Conv. Ctr. Room 15B

Advances in artificial intelligence are transforming microscopy from a tool of observation into a tool of discovery. Modern machine learning algorithms—especially deep neural networks—enable us to reconstruct, segment, and interpret microscopic images with unprecedented speed and precision. In this talk, I will discuss how AI can enhance every step of the microscopy pipeline, from data acquisition and noise reduction to virtual staining and quantitative analysis. Drawing on examples from my group’s work—including the DeepTrack framework for physics-aware deep learning and AI-driven virtual microscopy—I will show how we can train models to reveal hidden structures, track dynamic processes, and even uncover underlying physical laws directly from images. Beyond technical performance, I will also address the broader implications of this transformation: how AI reshapes the role of the experimenter, the reproducibility of quantitative microscopy, and the future of scientific understanding in an era of data-driven discovery.

EquiNET: A new technique for estimating Equilibrium free-energy differences from Non-Equilibrium Trajectories on arXiv

EquiNET uses a graph neural network to infer entropy production from non-equilibrium molecular trajectories and uses this information to estimate equilibrium free-energy differences. *(Image by S. K. Manikandan.)*

Free-Energy Differences from Nonequilibrium Fluctuations in High Dissipation
Sreekanth K Manikandan and Giovanni Volpe
arXiv: 2608.23394

Equilibrium free-energy differences can be measured from nonequilibrium work fluctuations using fluctuation theorems, such as the Jarzynski equality and the Crooks fluctuation theorem. However, these approaches become statistically inefficient in high-dissipation regimes because their convergence requires trajectories with negative entropy production, events that occur exponentially rarely. Here, we demonstrate that repeated measurements of trajectory fluctuations are sufficient to determine entropy production without relying on such rare events, or to determine a lower bound on it when only partial measurements are available. This information then yields exact estimates of free-energy differences, or rigorous bounds. We validate this approach, named EquiNET, with numerical simulations, including one of biomolecular folding and unfolding, showing that it recovers accurate free-energy estimates even in regimes where conventional approaches fail.

Poster presentation by G. Volpe at SPIE-ETAI, San Diego, 24 August 2026

An energy- and data-efficient foundation model for high-content phenotypic screening
Benjamin Midtvedt, Jesús D. Pineda Castro, Henrik Klein Moberg, Jim Denholm, Mattias Goksör, Christos Matsoukas, Bjarki Johannesson, Giovanni Volpe
Date: 24 August 2026
Time: 5:30 PM – 7:30 PM PDT
Place: Conv. Ctr. Exhibit Hall B1

High-content screening (HCS) enables large-scale characterization of cellular responses to chemical and genetic perturbations, yet current deep learning approaches remain limited by label scarcity, batch effects, and high computational cost. We present an energy- and data-efficient foundation model for high-content phenotypic screening that unifies these challenges within a single architecture. Trained on the full JUMP Cell Painting dataset, the model combines active learning with contrastive supervision to focus representation learning on biologically informative variation while reducing sensitivity to technical artifacts. By supporting multimodal integration of molecular metadata alongside fluorescence morphology, the model learns generalizable cellular representations that transfer across experimental conditions and downstream phenotypic tasks.

Presentation by A. Ciarlo at SPIE-OTOM, San Diego, 23 August 2026

Illustration of three different experiments autonomously performed by the SmartTrap system: DNA pulling experiments (top), red blood cell stretching (bottom left), and particle-particle interaction measurements (bottom right). *(Image by M. Selin.)*

SmartTrap: high-throughput optical tweezers via real-time adaptive control
Antonio Ciarlo, Martin Selin, Giuseppe Pesce, Lars Bengtsson, Joan Camunas-Soler, Vinoth Sundar Rajan, Fredrik Westerlund, L. Marcus Wilhelmsson, Isabel Pastor, Felix Ritort, Steven B. Smith, Carlos Bustamante, Giovanni Volpe
Date: 23 August 2026
Time: 4:55 PM – 5:10 PM PDT
Place: Conv. Ctr. Room 5B

We present SmartTrap, a fully autonomous optical tweezers platform that automates optical trapping experiments. SmartTrap uses deep learning–based 3D particle tracking, adaptive feedback control, and automated microfluidics to perform complete force spectroscopy and manipulation protocols without human intervention. Once initialized, the system operates continuously, autonomously handling trapping, molecular attachment, force application, data acquisition, and bead replacement. We demonstrate its capabilities using automated λ-DNA pulling experiments, which enable high-throughput force-extension and kinetic measurements. These capabilities are implemented using fully open-source control and analysis software with user-friendly interfaces, facilitating adaptation to existing optical tweezers setups and experimental workflows. SmartTrap is a flexible, general-purpose platform for automated optical trapping experiments in biomolecular, colloidal, and cellular systems.

Presentation by A. Lech at SPIE-ETAI, San Diego, 23 August 2026

deeplay: enhancing pytorch with customizable and reusable neural networks
Alex Lech, Mirja Granfors, Jiacheng Huang, Benjamin Midtvedt, Jesús Pineda, Harshith Bachimanchi, Carlo Manzo, Giovanni Volpe
Date: 23 August 2026
Time: 2:15 PM – 2:30 PM PDT
Place: Conv. Ctr. Room 2

Deeplay is a Python-based deep learning library that extends PyTorch, addressing limitations in modularity and reusability commonly encountered in neural network development. Built with a core philosophy of modularity and adaptability. Unlike traditional PyTorch modules, Deeplay allows the properties of submodules to be modified after creation. Layers, blocks, or activation functions can be replaced or reconfigured without rewriting the surrounding architecture, which reduces boilerplate code, improves reuse, and simplifies ablation studies.

Presentation by A. Ciarlo at SPIE-MNM, San Diego, 23 August 2026

Top: single gear; Bottom: the second gear from the right has an optical metamaterial that react to laserlight and makes the gear move. All gears are made in silica directly on a chip. Each gear is about 0.016 mm in diameter. *(Image by G. Wang)*

Light-driven micromechanical gear systems
Antonio Ciarlo, Gan Wang, Marcel Rey, Mohanmmad Mahdi Shanei, Kunli Xiong, Giuseppe Pesce, Mikael Käll and Giovanni Volpe
Date: 23 August 2026
Time: 9:30 AM – 9:45 AM PDT
Place: Conv. Ctr. Room 11A

We present microscopic geared metamachines driven by optical metasurfaces and fabricated using standard lithography techniques. Metasurface elements convert incident plane-wave illumination into localized optical forces and torques, enabling direct on-chip actuation of micromechanical components without external drives. We demonstrate microscopic gear trains powered by a single optically activated gear and a pinion-and-rack micromachine capable of rotational-to-linear motion conversion and controlled mirror actuation. The platform is compatible with planar fabrication and scalable to parallel integration, providing a compact approach to chip-integrated optical actuation and micromechanical functionality.

Keynote presentation by G. Volpe at SPIE-MNM, San Diego, 23 August 2026

Light-driven metamachines: From metarotors to integrated microscopic gear trains
(Keynote Presentation)
Giovanni Volpe
Date: 23 August 2026
Time: 8:30 AM – 9:05 AM PDT
Place: Conv. Ctr. Room 11A

Microscopic geared metamachines provide a powerful platform to control motion and mechanical work at the microscale. While the miniaturization of mechanical systems has been pursued for decades, integrating drives and gear trains at micrometer scales has remained a major challenge. Recent advances in optical metasurfaces and on-chip fabrication enable light-driven metarotors that transfer torque to passive gears under uniform illumination. These systems have been explored under various conditions by varying light polarization, intensity, and structural symmetry to achieve directional control, tunable rotation rates, and efficient torque transmission. Integrated designs such as rack-and-pinion configurations further demonstrate conversion of rotational motion into linear actuation within compact microsystems. More broadly, geared metamachines extend the landscape of light-powered micro- and nanomachines. In this presentation, I’ll give an overview of this field with a focus on metarotor design, torque transfer across microscopic gear trains, and integrated light-programmable micromechanical systems.

Global graph features unveiled by unsupervised deep learning published in Machine Learning: Science and Technology

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 deep learning
Mirja Granfors, Jesús Pineda, Blanca Zufiria Gerbolés, Joana B. Pereira, Carlo Manzo, Giovanni Volpe
Machine Learning: Science and Technology, 7, 045054 (2026)
arXiv: 2503.05560
doi: 10.1088/2632-2153/ae8d7f

Graphs provide a powerful framework for modeling complex systems, but their structural variability poses significant challenges for analysis and classification. To address these challenges, we introduce Graph Autoencoder Uncovering Descriptive Information (GAUDI), an unsupervised graph deep learning framework designed to capture both local details and global structure. GAUDI employs an hourglass architecture with hierarchical pooling and upsampling layers linked through skip connections, which preserve essential connectivity information throughout the encoding–decoding process. Even though identical or highly similar underlying parameters describing a system’s state can lead to significant variability in graph realizations, GAUDI consistently maps them into nearby regions of a structured and continuous latent space, effectively disentangling invariant process-level features from stochastic noise. We demonstrate GAUDI’s versatility across multiple applications, including small-world networks modeling, characterization of protein assemblies from super-resolution microscopy, analysis of collective motion in the Vicsek model, and identification of age-related changes in brain connectivity. Comparison with related approaches highlights GAUDI’s superior performance in analyzing complex graphs, providing new insights into emergent phenomena across diverse scientific domains.