About the project

Objective

The overall goal of the project is to develop optimization frameworks to assist in process design, learning accurate models from process data, and support optimal decision-making. The optimization algorithms and frameworks will be tailored towards the needs and processes of interest for LKAB. The type of optimization problems considered in the project falls into the category of “Gray-box” optimization, where parts of the optimization problem are known analytically, but some parts are given by a simulator. The simulator is not necessarily a complete black box, but the simulation model can be too complex to be integrated directly into the optimization model.

Designing industrial processes is a challenging task and often requires the use of advanced design and simulation software. For example, accurately describing the process and physical phenomena involved often requires the solution of systems of partial differential equations (PDEs) or a Computational Fluid Dynamics (CFD) simulation. Therefore, evaluating the impact of simply changing a single design parameter can require running a CFD or solving a large system of PDEs and may also require access to chemical and thermodynamic libraries. Solving such systems is a challenge on its own, and in practice, requires the use of advanced simulation software. The main downside of such software is the inherent black-box nature; you can evaluate a single specific design choice at a time, but the end-user does not gain more knowledge. In practice, when such software is used for process design the engineers often follow some common rules of thumb and trial and error to come up with a good design that can be evaluated by software and later put into production. However, the resulting design might be far from globally optimal, i.e., there might exist a far more superior design. Employing such a sub-optimal solution can, e.g., result in increased environmental impact due to a higher energy consumption, increased raw material usage and waste. Furthermore, when used in investment planning and feasibility studies, the sub-optimal designs can cause superior technologies/solutions to seem unreasonably expensive and maybe even economically infeasible. Therefore, there is a strong need for frameworks that enable the combination of simulation software with optimization algorithms to find optimal process designs. 

Background

LKAB is an international mining and minerals group that supplies sustainable iron ore, minerals and specialty products. Since 1890, the company has evolved through unique innovations and technology solutions and is driven forward by more than 4,500 employees in 12 countries. The company is the largest supplier of iron ore in the European Union and a key player in the transformation of the iron and steel industry towards sustainability. LKAB’s goal is to develop carbon-free processes and products by 2045.

The work to eliminate carbon emissions creates new challenges and opportunities. Potential routes and new ideas are continuously being investigated and evaluated. As the production processes are complex and experiments often expensive and time-consuming, the feasibility and potential of new alternatives are often investigated numerically through computer simulations. Currently, computer simulations are used in a black-box fashion, i.e., you can evaluate a specific configuration of process parameters, but you don’t gain any more information or knowledge of the best configurations. Therefore, it is of utmost importance to utilize the full potential of the models and to find the best solutions. This project aims to provide a key component for a systematic model-based design approach.

Increasing computational power, increasing amount of data, and overall digitalization are continuously boosting the potential benefits of extensive computational simulations in process design, but how to best utilize simulators is not trivial. To learn and extract knowledge from the process simulations, the project aims to develop deterministic optimization methods and frameworks for determining optimal designs and operations.

Crossdisciplinary collaboration

The project is a collaboration between LKAB and KTH, where we are focusing on developing optimization algorithms suitable for tackling challenging optimization tasks in industrial production processes. The project brings together expert knowledge from Process Systems Engineering, Applied Mathematics, Optimization, and Process Simulation.

Participating in the project:

About the project

Objective
This project aims to develop and evaluate algorithms for dynamic inference offloading in Scania’s fleet, optimizing the trade-off between model accuracy, latency, and resource consumption. The proposed cascaded inference approach ensures that data is first processed by small models at the edge (M1). Depending on inference confidence, decisions are made in real-time on whether to offload tasks to more complex models (M2-M4) with higher accuracy but increased computational costs.

The core objectives include: (1) designing algorithms that improve inference accuracy while minimizing latency and bandwidth usage; (2) providing theoretical guarantees for the proposed methods, particularly in terms of regret minimization; and (3) benchmarking these algorithms using real-world data from Scania’s operational vehicles. The research will contribute to optimizing deep learning model deployment, ensuring scalable and efficient AI integration in industrial applications while maintaining cost-effectiveness and system reliability.

Background
Deep Learning (DL) models have become a standard for data-driven tasks such as classification and predictive analytics due to their high accuracy. However, their computational and memory demands often require cloud-based deployment, which introduces challenges like latency, bandwidth consumption, and security concerns. In response, edge computing has gained traction, enabling inference on resource-constrained Edge Devices (EDs) such as IoT sensors, mobile devices, and autonomous vehicles. While edge deployment reduces communication delays and enhances data privacy, small models often suffer from lower accuracy.

Scania, a global manufacturer of commercial vehicles, faces similar challenges in deploying DL models across its fleet for tasks such as autonomous driving, predictive maintenance, and sustainable operations. There exists a trade-off between accuracy and efficiency when placing models at different computation points. This project explores cascaded inference, where small models operate locally, and only complex cases are offloaded to more powerful computing resources, balancing accuracy with cost efficiency.

Partner Postdocs
This project brings together experts from multiple disciplines to address the challenges of deploying DL models efficiently across Scania’s fleet. Researchers from machine learning, optimization, embedded systems, and automotive engineering will collaborate to develop cascaded inference strategies that optimize accuracy, latency, and resource usage.

Scania’s senior data scientists, Dr. Sophia Zhang Pettersson and Dr. Kuo-Yun Liang, provide real-world insights into vehicle data, predictive maintenance, and cost modelling. Associate Prof. Lei Feng adds knowledge in Bayesian optimization and deep learning techniques for edge computing.

This collaboration ensures that theoretical advancements in machine learning align with practical deployment challenges in commercial vehicles. By integrating perspectives from academia and industry, the project fosters innovation in scalable AI solutions, leading to efficient, adaptive, and cost-effective DL deployment across connected fleets.

Supervisor
KTH researchers, led by Prof. James Gross, contribute expertise in hierarchical inference, algorithm development, and performance guarantees.

About the project

Objective
This project will explore the neural correlates of human-human and human-robot conversations, with the goal of creating adaptive social robots capable of fostering more meaningful interactions. Social robots can assist people in societal situations such as health care, elderly care, education, public spaces and homes.

Our newly developed telepresence system for human-robot interaction, allowed participants to situate themselves in natural conversations while physically in a functional magnetic resonance imaging scanner. Each participant interacted directly with a human-like robot or ahuman actor while lying in the scanner. In our previous research pairs project, we used this telepresence interface to create the pioneering NeuroEngage fMRI dataset.

This project aims to advance the understanding of conversational engagement by integrating neuroscience, human-robot interaction (HRI), and artificial intelligence. Engagement plays a crucial role in effective communication, yet its underlying brain mechanisms and real-time detection remain largely unexplored. We will use the NeuroEngage dataset and complement it with additional multimodal features like facial expressions, audio embeddings, and detailed annotations of engagement levels. By using multimodal machine learning (MML), this research will develop models capable of detecting and responding to engagement levels in social interactions.

Background
In everyday conversations, a speaker and a listener are involved in a common project that relies on close coordination, requiring each participant’s continuous attention and related engagement. However, current engagement detection methods lack robustness and often rely on superficial behavioral cues, without considering the underlying neural mechanisms that drive engagement. Prior research has demonstrated the feasibility of engagement detection using multimodal signals, but most existing datasets are limited in their scope and do not incorporate neuroimaging data.

In our previous work, by analyzing two different datasets, we have shown that listening to arobot recruits more activity in sensory regions, including auditory and visual areas. We also have observed strong indications that speaking to a human, compared to the robot, recruitsmore activity in frontal regions associated with socio-pragmatic processing, i.e. considering the other’s viewpoint and factoring in what to say next. Additional comparisons of this sort will be enabled by expanding our dataset and refining machine learning models for engagement prediction. As a result, this project will help with AI-driven conversational adaptivity, advancing research in both HRI and neuroscience.

Crossdisciplinary collaboration
The researchers in the team represent the Department of Intelligent Systems, division of Speech Music and Interaction at KTH EECS, the Psychology Department and the Linguistics Department at Stockholm University. This project integrates neuroscience, linguistics, social robotics, and AI to study how humans engage in conversations with both humans and robots.

About the project

Objective
Medical doctors often face difficulty to choose a set of medicines from many options available for a patient. Medication is expected to be disease-specific as well as person-specific. Individual patients may respond differently to the same medication, so the selection of medication should be personalized to everyone’s needs and medical history. In this project, we will explore how AI (artificial intelligence) can help doctors to identify existing medications and/or therapies that can be repurposed for the treatment of dementia.

Dementia is a large-scale health care problem, where around 10% of the population more than 65 years of age suffers from it. Therefore, if AI can assist clinicians in medication selection for dementia patients, it would lead to a significant improvement in the efficiency of treatment. AI can also predict decline (or worsening) of a patient’s health condition over time. Clinicians and healthcare systems will then get precious time to decide life-saving interventions. This heralds use of AI-based medications. The pressing question: can we trust AI systems, mainly its core called machine learning (ML) for patient data analysis and predictions to doctors? Can the ML algorithms explain their predictions to the doctors? In a joint collaboration with Karolinska University Hospital (KUH), Karolinska Institute (KI) and KTH, we will develop trust-worthy ML algorithms with explainable results, and then explore the algorithms to discover new uses for approved medications that were originally developed for different medical conditions.

Background
XML based medication repurposing for dementia (XMLD) refers to the development and application of XML algorithms to identify potential drugs among existing drugs or medications that can be repurposed for the treatment or management of dementia. The goal is to develop XML algorithms to discover new uses for approved drugs or therapies that were originally developed for different medical conditions in patients with dementia. Therefore, for a patient and/or a class of patients, identification of potential drugs among many existing drugs is a variable selection problem, where XML can help.

Therefore, The XML algorithms will be developed to analyze and identify patterns, relationships, and potential associations between drug characteristics, disease severity, and patient outcomes. There are many advantages of medication repurposing for dementia using XML, such as cost and time saving, safety profile, broad range of medication candidates, and improved treatment efficiency. Overall, it addresses a pressing healthcare problem with potentially widespread impact. While our focus is dementia in this project, the accumulated technological knowledge can be used for medication repurposing of many other health problems and diseases in clinics. The proposed XMLD project will establish a strong cooperation between medical doctors and ML researchers in the clinical environment.

Partner Postdoc
Xinqi Bao

Main supervisor
Saikat Chatterjee

Co-supervisor(s)
Martina Scolamiero

About the project

Objective
The SENZ-Lab project develops and validates a cost-efficient, real-time, dynamic sparse sensing approach for urban traffic monitoring and environmental footprint assessment in Stockholm’s Environmental Zone Class 3. Using acoustic sensors and AI-driven modelling, it seeks to establish a 2D digital twin of the city’s traffic, enabling real-time monitoring of noise, air pollution, and vehicle-level activity. The goal is to enhance traffic management, reduce emissions, and support sustainable urban mobility.

Background
As cities expand, noise and air pollution pose significant health risks. Traditional monitoring methods struggle with real-world complexity, requiring new solutions. Building on previous research in Stockholm’s Hornsgatan innovation zone, this project integrates IoT, AI, and real-time traffic simulations to improve monitoring accuracy and inform urban policy. The initiative aligns with Stockholm’s environmental goals and KTH’s strategic pillars of sustainability and digitalization.

Crossdisciplinary collaboration
The project brings together experts from multiple fields, including urban sensing, traffic modeling, AI, and GIS-based visualization. The consortium includes two research teams at KTH specialized in Acoustics and Geoinformatics, supported by the City of Stockholm, combining academic research with real-world urban planning needs. By integrating cutting-edge technology with policy-driven insights, the project provides practical solutions for creating quieter, healthier, and more sustainable cities.

About the project

Objective
We develop a novel multimodal imaging database, PelvicMIM, by integrating next-generation digital diagnostic technologies to advance the evaluation of childbirth-related pelvic floor muscle injuries. This effort includes the development and validation of cutting-edge imaging modalities—Shear Wave Elastography (SWE), Magnetic Resonance Elastography (MRE), and Diffusion Tensor Imaging (DTI). These techniques will be applied in vivo to quantify the biomechanical and structural properties of pelvic floor muscles. A deep learning-based image processing framework will be designed for multimodal image registration, enabling the overlay of stiffness maps from MRE/SWE and fiber orientations from DTI onto MRI and ultrasound images. Our proposed approach facilitates cross-modality findings, offering deeper insights into muscle function and injury mechanisms.

Background
One in two middle-aged women suffer from pelvic floor dysfunction such as urinary and fecal incontinence or prolapse of the pelvic organs into the vagina, which profoundly impair quality of life. Injuries to the pelvic floor muscles due to birth are highly associated with pelvic floor dysfunction later in life. Nevertheless, injuries to these muscles, which cannot be surgically repaired, have been largely ignored and poorly studied. The Swedish Agency for Health Technology Assessment, SBU, has identified birth-related injuries to the levator ani muscle (LAM), the three largest muscles of the pelvis, as a priority area for research (April 2019). Although recent research also highlights the urgent need for quantitative assessment of LAM injuries, clinical practice still relies on conventional ultrasound, which lacks the ability to quantify biomechanical or structural properties that are important indicators of soft tissue health. These properties are crucial for the assessment of the LAM, as it is a complex structure of three muscles working together in a sheet-like shape with different layers and fiber directions.

Crossdisciplinary collaboration
The team of researchers is composed of members from the KTH School of Engineering Sciences in Chemistry, Biotechnology and Health, Department of Biomedical Engineering and Health Systems and KTH School of Engineering Science, Department of Engineering Mechanics. The project is conducted in close collaboration with clinical partners at Karolinska University Hospital.

About the project

Objective
The project envisions a mobile cyber-physical system where people carrying mobile sensors (e.g., smartphones, smartcards) generate large amounts of trajectory data that is used to sense and monitor human interactions with physical and social environments. Built upon the static causal inference results in the cAIMBER project, the CIML4MOB project aims to build causally informed machine learning models for predicting adoption time of individuals and subpopulations and their risks of attrition by input dates. Such dynamic causal models may then drive policy design strategies for lasting behavioral changes (the ultimate purpose of behavior interventions).

Background
The ever-changing mobility landscape and climate change continue to challenge existing operating models and the responsiveness of city planners, policymakers, and regulators. City authorities have growing investment needs that require more focused operations and management strategies that align mobility portfolios to societal goals. The project targets the root cause of traffic (human) and proposes causally informed machine learning to learn and predict human mobility dynamics from pervasive mobile sensing data that helps cities meet both sustainability challenges and improve urban resilience to disruptive events.

The human mobility dynamic problem is defined to predict travel choice decisions given a set of factors, including for example individual traits, travel contexts, and interventions. The research pair project (cAIMBER, 2022-2024) developed the data-driven causal inference method to discover the static causal graph of behavior responses to interventions in public transport. The cAIMBER causal model allows for analysis and prediction of human behavior based on population features, but without regard to when individuals or other subpopulations will adopt the desired behavior of a certain incentivization program. From the perspective of city planning and utility costs, two fundamental questions are (1) how to incentivize early adoption of the desired behavioral shift (adoption time) and (2) given an individual has shifted their behavior, how to prevent reversion to baseline behavior (attrition time). The research consolidator project, CIML4MOB, aims to build upon cAIMBER results to build causally informed machine learning models for predicting adoption time of individuals and subpopulations and their risks of attrition by input dates.

Crossdisciplinary collaboration
The research collaborates between researchers in transportation science and mathematics at KTH.