About the project
Objective
The aim is to establish a mathematical and computational framework for gradient-flow variational inference over families of probabilistic circuits.
Gradient flows in the Bures–Wasserstein and Wasserstein–Fisher–Rao (Hellinger–Kantorovich) geometries are well developed, with particularly explicit results for Gaussian distributions. However, it remains an open question how these frameworks extend to more expressive model classes such as probabilistic circuits, which generalise many tractable models while still supporting exact and efficient inference. In this project, we will investigate how Bures–Wasserstein and Wasserstein–Fisher–Rao gradient flows can be extended to probabilistic circuits. This will yield principled, geometry-aware algorithms for fitting expressive yet tractable variational approximations, with applications in robotics, AI for science, and computational epidemiology.
Background
Approximating intractable distributions, such as Bayesian posteriors, with an optimised tractable surrogate through variational inference is a cornerstone of probabilistic machine learning, with applications in robotics, AI for Science, healthcare, and computational epidemiology. Despite its popularity, modern variational inference faces a dilemma: mathematically rigorous frameworks exist, but they are limited to restrictive families of surrogates, such as Gaussians, which in turn limits approximation quality.
In parallel, probabilistic circuits have found increasing attention in the tractable probabilistic modelling community as a unifying framework. They offer a promising avenue to represent richer variational families, as they can represent complex, multimodal distributions while guaranteeing exact, polynomial-time density evaluation, marginalisation and moment computations. However, the geometry of families of probabilistic circuits within Wasserstein space is largely unexplored, and how this geometry interacts with structural conditions which enable tractable computations is not yet understood.
This project unites these two complementary research directions to develop mathematically rigorous and computationally effective gradient-flow variational inference techniques over families of probabilistic circuits.
Crossdisciplinary collaboration
The project established a collaboration between Trapp’s group in the Computational and Learning Systems department and Zhu’s group in the Mathematics department at KTH Royal Institute of Technology and brings together expertise on tractable probabilistic modelling with probabilistic circuits (Trapp) and mathematical foundations on gradient-flow variational inference (Zhu).
About the project
Objective
This project aims to develop autonomous AI scientist systems for bioinformatics analysis of spatial transcriptomics data. Building on general-purpose AI models, these systems will independently write and execute code, critically evaluate results, refine analytical methods, and generate biological hypotheses about cellular architecture of human tissues. We will compare single-agent and multiagent approaches and establish methods for evaluating accuracy and reliability under human expert supervision. By applying these systems across gastrointestinal diseases, we aim to accelerate biomedical research in mucosal immunology, exploring how AI advances can contribute to scientific discovery in the life sciences.
Background
Spatial transcriptomics map gene expression and cellular interactions within human tissues, providing unprecedented insights into health and disease. However, rigorous data analysis requires extensive computational expertise and workflow optimization to curate data and accurately model cellular behaviour, which limits large-scale analysis of publicly available spatial transcriptomic data. In contrast, meta-analysis of single-cell gene expression data has been caried out at scale by the Human Cell Atlas consortium across many human organs, but studies of spatial transcriptomics data are lacking.
Recent advances in AI offer opportunities to automate scientific data analysis and discovery. To test this, the project will use 41 publicly available spatial transcriptomics datasets from healthy intestinal tissues, inflammatory bowel disease, and colorectal cancer to benchmark autonomous AI scientists. The tasks include cell type annotation, niche identification, and cross-disease integration, modeling cell states, niches, and technical biases from diverse spatial transcriptomic platforms. Beyond identifying cell types and tissue structures, the project will investigate whether AI can generate novel biological predictions about immune cell migration, which will be tested experimentally with automated cytokine perturbations.
Crossdisciplinary collaboration
The project establishes a cross-departmental collaboration between the labs of Simon Koplev and Wei Ouyang, combining expertise in computational biology, mucosal immunology, spatial transcriptomics, agentic AI, and automated microscopy. Koplev’s group will lead the biological data integration, benchmarking, and interpretation, while Ouyang’s group will develop autonomous AI systems and experimental validation workflows. Together, the research will establish AI-assisted biomedical discovery of cells and niches in the human gut, connecting computational analysis with automated biological experimentation. Concretely, the successful establishment of a reference spatial transcriptomic dataset for the human gut would represent a foundational data resource for the gastrointestinal research community, promoting cross-disciplinary inquiry across studies and gastrointestinal diseases.
About the project
Objective
AURAS aims to develop and validate a catheter-deployable intravascular digital sensing platform for continuous monitoring of valvular function through synchronized acoustic and ECG signals. The project will first develop an intravascular sensing unit capable of acquiring high-fidelity acoustic and ECG signals within the cardiovascular system and characterize its performance under physiologically relevant conditions. It will combine advanced signal processing, multimodal data analysis and machine learning to identify novel digital biomarkers of valve dysfunction and characterize different pathological conditions. Finally, the integrated technology will be validated in large-animal models against multimodal clinical reference measurements, assessing its feasibility, robustness and potential for early detection of valvular dysfunction.
Background
Advances in congenital heart disease care have led to a rapidly growing population of adults living with lifelong cardiovascular conditions that require continuous surveillance. In particular, patients with congenital or early-onset valvular disease may experience progressive deterioration of valve function, including stenosis, regurgitation, leaflet stiffening and calcification. This deterioration may progress silently, without noticeable symptoms, until significant cardiac damage has occurred. Current clinical follow-up is largely based on periodic hospital assessments, including imaging and surface auscultation. However, these episodic measurements provide only snapshots of cardiac function and may miss early mechanical changes occurring between clinical visits. Existing implantable monitoring technologies predominantly focus on electrical activity and therefore provide limited information about the mechanical behaviour of heart valves.
AURAS addresses this gap by exploring a new approach to continuous cardiovascular monitoring. The project will combine high-fidelity intravascular acoustic sensing with synchronized ECG measurements to capture both the electrical and mechanical aspects of cardiac function. Advanced signal processing and machine learning will then be used to identify digital signatures associated with valve function and dysfunction. By moving from episodic assessment towards continuous, data-driven monitoring, AURAS aims to establish the technological foundations for earlier detection of silent valve deterioration and more personalized cardiovascular surveillance.
About the Digital Futures Postdoc Fellow
Federica Semproni is a biomedical engineer and postdoctoral researcher at the Department of Biomedical Engineering and Health Systems at KTH. She received her B.Sc. and M.Sc. in Biomedical Engineering from Politecnico di Torino and her Ph.D. in BioRobotics from Scuola Superiore Sant’Anna, all with honours.
Her research combines biomedical engineering, soft robotics, implantable technologies, and data-driven approaches to develop minimally invasive systems for physiological monitoring and intervention. During her M.Sc., she conducted research at Harvard Medical School, where she developed expertise in artificial intelligence and machine learning. Her Ph.D. focused on soft robotic technologies for artificial organs, with particular emphasis on implantable sensing and actuation systems for artificial urinary bladders, from prototyping to in vivo validation. As a visiting Ph.D. researcher at Stanford University, she further developed her expertise in cardiovascular applications, working on implantable sensing and electrohydraulic actuation for cardiovascular monitoring and circulatory support.
At KTH, she brings these complementary experiences together to develop implantable technologies for continuous cardiac valve monitoring, bridging engineering innovation, AI-driven data analysis, and clinical needs.
Main supervisor
Seraina Dual, KTH Royal Institute of Technology.
Co-supervisors
Frank Niklaus, KTH Royal Institute of Technology.
Erica Zeglio, Stockholm University.
About the project
Objective
UrbanEI aims to advance AI-powered Earth Observation (EO) for urban environmental intelligence by mapping and monitoring urban green infrastructure (UGI), assessing urban heat hazards, and detecting illegal waste dumping and burning. The project will integrate satellite EO, environmental, and crowd-sourced data within geospatial foundation models and generative AI frameworks. EO foundation models will be adapted to process large-scale multimodal datasets, enabling detailed UGI mapping, heat-hazard detection, and improved urban heat-risk assessment.
Generative AI will be explored to produce super-resolution imagery for high-resolution UGI mapping and daily land surface temperature estimation, and to generate city-wide air-temperature surfaces from sparse crowd-sourced measurements. In parallel, UrbanEI will pioneer the use of generative AI and open EO data, including Sentinel-2 imagery, for early detection and monitoring of illegal waste dumping and burning.
Expected outcomes include dynamic high-resolution maps of urban green infrastructure, improved city-scale air-temperature and urban heat-island assessments, and earlier detection of illegal waste activities. UrbanEI will advance the state of the art in AI and EO for urban environmental intelligence and provide timely, reliable, and actionable information for climate adaptation, environmental monitoring, and sustainable urban development.
Background
With 2024 marking the warmest year on record and 2026 ranking as the second-warmest year to date, the accelerating impacts of climate change are increasingly evident. Intensifying heatwaves and extreme weather pose major challenges, particularly in rapidly growing cities where heat, environmental degradation, and illegal waste activities threaten public health, infrastructure, and sustainable development. There is an urgent need for timely, accurate, and actionable environmental intelligence to support urban monitoring, mitigation, and adaptation.
Cross-disciplinary collaboration
To be announced
PIs of the project
Yifang Ban, KTH
Josephine Sullivan, KTH
Sebastian Hafner, RISE
About the project
Objective
This project will develop efficient representation and learning methods for system dynamics that allow a robot to compensate for the inference delay of large robot action models, e.g., vision-language-action (VLA) models. The robot acquires and refines a predictive modelthrough its interaction with the environment. Through rolling out the learned system dynamics and generating the corresponding observations, the policy can be conditioned on ananticipated observation rather than an outdated one.
Background
VLA models allow robots to accomplish a wide range of tasks and have attracted intense interest in both academia and industry. Their capability comes at the expense of long inference time. By the moment an action or action chunk is computed, the robot state and the world have already evolved, and the robot acts on an observation of the past. Closing this gap requires anticipating the state at execution time and the observation that corresponds to it. For the robot’s own state, the prediction is straightforward. In dynamic scenes where objects roll, slide or deform in respond to robot actions, it becomes considerably harder.
About the Digital Futures Postdoc Fellow
Yunfan Gao completed her Ph.D. in 2026 at the University of Freiburg under the supervision of Prof. Moritz Diehl, and simultaneously, until September 2025 she was an industrial Ph.D.student at Bosch Center for AI, under the industrial supervision of Dr. Niels van Duijkeren. Previously, she obtained her master’s degree in Robotics, Systems and Control from ETH Zurich in 2022 and her bachelor’s degree in Electronic Engineering from Fudan University in 2019.
Main supervisor
Danica Kragic Jensfelt, KTH Royal Institute of Technology.
Co-supervisor
Florian Pokorny, KTH Royal Institute of Technology.
About the project
Objective
The overall objective of this project is to promote deeper, sustained research engagement between Digital Futures members, both within the Learn Working Group and within Digital Futures as a whole. The Research Sparks program accomplishes this by introducing a framework for fostering long-term collaboration and knowledge exchange that solidifies the working group structures as attractive and co-dependent research environments. Research Sparks will stimulate long-term engagement through a careful combination of guided activities that will help researchers demystify their research for others, encourage hands-on and interactive learning, boost research productivity and increase the visibility of Digital Futures research activities.
Background
The project was motivated by the feedback obtained from participants in the Learn Working Group’s 2026 kick-off event on 3 March 2026. The participants requested more activities that deepened their understanding of the research of other Digital Futures members and their newfound connections, with an eye towards opportunities for organized research collaborations.
Cross-disciplinary collaboration
The Research Sparks program is specifically designed to stimulate cross-disciplinary collaboration. It is a semester long program that begins with a series of kick-off lectures that “demystify” a research topic of active focus within the Digital Futures community. These lectures aim to answer the questions, “What is this field and its driving goals? Why are there researchers at Digital Futures studying it? How does it connect to my own research?” Following these lectures, interested researchers will have the opportunity to participate in hands-on learning a
ctivities to deepen their understanding of the topic, as well as to participate in research collaborations run by experts in the topic. These ongoing activities will be supported by seminars and workshops that help the participants showcase their work and connect with international experts on the topic. Participants will learn about the new field through their engagement in these research activities, which help them broaden their own research and connect it with a research topic at the heart of Digital Futures activities.
PI: Liam Solus
Co-PIs: Kathlén Kohn, Iolanda Leite, Henrik Boström, Sebastian Dalleiger
About the project
Objective
Cancer progression is driven not only by genetic alterations within individual cells but also by complex communication between cancer cells and their surrounding microenvironment. Understanding and controlling these interactions would enable therapies that disrupt cancer-promoting cellular equilibria, improve treatment design, and accelerate precision medicine. The major technical challenge is to develop predictive computational models that capture high-dimensional, nonlinear cell-cell interactions and reliably forecast the effects of targeted interventions. A promising path to address this challenge is to combine biological knowledge on cellular signaling and regulatory mechanisms with recent advances in AI and dynamical systems theory. This would entail learning biologically consistent models of multicellular behavior, validated and enriched on micro-physiological systems.
As a first approach, we focus on communication between pancreatic cancer cells and stromal cells, whose reciprocal interactions are known to reinforce tumor growth and therapeutic resistance. While recent single-cell technologies provide unprecedented measurements of cellular states and signaling pathways, these data remain largely observational and therefore have limited predictive power for designing interventions. Moreover, existing AI approaches rarely account for the coupled dynamics governing cellular communication, making the identification of stable and controllable cellular states particularly challenging. This motivates our project, SCALE, to develop biology-informed AI methods for modeling and controlling multicellular systems. The objectives of SCALE are to establish the fundamental theory and computational tools needed to learn predictive models of cell-cell interactions, to identify healthy and disease-associated equilibria, and to design optimal in silico interventions to guide experimental validation and future therapeutic development.
Background
Cells continuously exchange signaling molecules to create the cellular microenvironment. In pancreatic cancer, communication between cancer cells and surrounding stromal cells regulates processes such as tumor growth and therapeutic resistance. In reality, these cell-cell interactions are not directly observed. Instead, they are studied through molecular measurements and cell culture experiments that capture the cellular states resulting from these interactions. Therefore, it is difficult to directly characterize the interaction mechanisms and predict how perturbations of one cell population influence the others.
The functional relationship between cell-cell interactions and the observed cellular states is governed by complex biological processes involving intracellular signaling and intercellular communication. For multicellular systems such as the pancreatic tumor microenvironment, these coupled dynamics remain difficult to characterize because of the large number of interacting molecules and feedback mechanisms. Furthermore, experimental observations are inherently noisy and capture only part of the underlying biological processes.
Naturally, modeling and predicting the coupled dynamics that govern cell-cell interactions from experimental observations constitute a challenging systems modeling problem.
Cross-disciplinary collaboration
The project will combine methods and techniques from separate research fields – (a) biological knowledge about virtual cell models and computational biology, (b) dynamical system and control theories, (c) micro physiological systems design, and experiments.