Pitch Sessions
Pitch sessions are organized in the @ Chapter House. The main dynamic pitch block takes place from 15:15 to 15:55. Get inspired by lightning-fast 2-minute presentations of FEA research.
Predictive Maintenance in Residential Heat Pump Systems
TimeAuthorsAbstractHeat pumps are a key technology in the transition towards energy-efficient residential heating and cooling, yet faults and performance degradation can undermine their efficiency, reliability and economic benefits. Heating, ventilation and air conditioning (HVAC) systems account for 38% of total energy consumption in buildings, equivalent to 12% of total energy consumption worldwide. A recent field study involving 1,023 residential heat pumps across Central Europe found that approximately 17% of air-source heat pumps failed to meet European efficiency standards. Undetected faults can lead to prolonged energy losses, reduced thermal comfort, unexpected breakdowns and costly repairs. Early fault detection is therefore essential for shifting from reactive maintenance towards proactive, condition-based monitoring. This research presents a data-driven anomaly detection framework for the early detection of emerging faults in residential air-to-air heat pumps. The framework addresses key challenges in residential environments, including limited sensor availability, diverse operating conditions, varying household usage patterns and the scarcity of labelled fault data. A self-supervised machine learning approach based on autoencoders is employed to learn normal system behaviour from operational data and identify deviations that may indicate developing faults. The framework investigates multiple fault types, including refrigerant leakage, compressor-related faults, piping malfunctions, and fan motor malfunctions, to assess its ability to detect different fault signatures under real-world operating conditions. It is developed and evaluated using real-world data collected from multiple households across different geographical regions, providing a challenging setting for assessing its applicability beyond controlled laboratory environments. The results demonstrate the potential of the proposed framework to identify early indications of faults, achieving an average warning lead time of 15 days before recorded fault confirmation. The proposed framework contributes towards more reliable predictive maintenance strategies for residential heat pumps, with the potential to reduce unexpected failures, improve maintenance planning and support energy-efficient building operation.
Read moreView location in Venue Guide →In-silico hemodynamic model for the assessment of mitral regurgitation quantification methods
TimeAuthorsAbstractIntroduction: Mitral valve regurgitation (MR) is the most common valvular disease, characterised by incomplete closure of the mitral valve. Leak severity can be quantified by the effective regurgitant orifice area (EROA) and regurgitant volume (RVol), which can be assessed using echocardiographic methods. However, current guidelines recommend confirming these findings using an integrative approach. The aim of our work is to develop an in-silico framework to enhance ultrasound-based MR quantification. Material and Methods: A geometrically idealized 3D computational fluid dynamics (CFD) model of the left heart was created, based on average adult dimensions. The mitral valve was modelled as a time-varying truncated cone, with opening and closing times set to 50 ms. To simulate a mild and severe MR case, the leakage area was set to 20 and 43 mm², respectively. The mesh consisted of about 500k tetrahedral elements and was remeshed when the skewness of elements exceeded 0.95 and the orthogonality deceeded 0.1. Four cardiac cycles of 1s were simulated in ANSYS Fluent, assuming laminar incompressible flow. In addition, the Proximal Isovelocity Surface (PISA) Method was evaluated by projecting the velocity field onto the beam direction and measuring the PISA radius. Results: Results show a large central jet forming in the left atrium during systole, reaching peak velocities of about 4.70 m/s for both cases. The RVol reached 29.82 ml and 63.06 ml for mild and severe MR, respectively. For severe regurgitation, a PISA radius of 0.72 cm was measured, resulting in an RVol of 52.91 ml. Conclusion: Under the imposed boundary conditions, mild and severe MR show similar flow fields and velocities, but a larger leakage area increases backward flow into the left atrium. In case of severe MR, PISA underestimates the RVol. Future work will incorporate more realistic anatomy and boundary conditions to better represent severe MR.
Read moreView location in Venue Guide →Model-based Efficiency and Lifetime Optimization of Drivetrains
TimeAuthorsAbstractThe efficiency and lifetime of drivetrains are key drivers for operational and maintenance costs, but also for the carbon footprint and, thereby, the sustainability of machinery. The lifetime of drivetrains is often determined by friction, wear, and fatigue behaviour of their critical components such as bearings, gears, and seals. This friction, wear, and fatigue behaviour can be predicted during the design phase using high-fidelity multi-body simulation models. However, open challenges remain. Firstly, efficiency and lifetime are affected by continuous wear- and fatigue induced damage in the contacts of tribological systems. Secondly, the high-fidelity simulations require significant computational resources, thereby limiting the number of simulations that can be conducted. Therefore, these models cannot be used efficiently for design purposes. As a result, the aim of this research is to develop high-fidelity models that can be used for optimization tasks and for computationally efficient lifetime predictions. First, a multi-body simulation model with elastic bodies and elastohydrodynamic coupling will be developed for a gearbox. This model will then be extended to account for transient wear- and fatigue-induced degradation of machine component. Lastly, machine-learning-based surrogate models will be developed to enable rapid computation of efficiency and lifetime at time instant, accounting for instantaneous and past operating conditions.
Read moreView location in Venue Guide →Catching Valuable Phenolics from Wood: A Computational Screening Framework
TimeAuthorsAbstractLignocellulosic biomass, which includes wood, straw and other plant materials, consists of three main parts, namely cellulose, hemicellulose, and lignin. Cellulose and hemicellulose make up the carbohydrate fraction, while lignin is a complex aromatic polymer that provides structural support and rigidity to plants. This polymer can be broken down into valuable small molecules such as vanillin, guaiacol and syringaldehyde. These molecules are used in food flavours, fragrances, pharmaceuticals and as building blocks for bio-based plastics, and today they are largely made from fossil raw materials. Obtaining them from biomass takes three steps: separating lignin from the carbohydrate fraction, depolymerisation (breaking the large lignin polymer into small molecules, for example by oxidation or hydrogenation), and separating these molecules from the resulting complex liquor that also contains salts, solvents, and residual sugars. This last step remains a bottleneck because the target molecules are present at low concentrations in a complex mixture, and several of them have similar chemical properties. The main separation strategies, such as liquid–liquid extraction, distillation and membrane filtration, can require large amounts of solvent or energy, or give limited selectivity. Adsorption, where molecules stick to the surface of a porous solid (a sorbent), is an interesting alternative because it works under mild conditions, can be selective, and allows the sorbent to be reused. Adsorption, where molecules stick to the surface of a porous solid (a sorbent), is attractive because it works under mild conditions, can be selective and allows the sorbent to be reused. We surveyed the literature on sorbents for lignin-derived phenolics: polymeric resins, zeolites, metal-organic frameworks (MOFs), covalent organic frameworks (COFs) and others. In the studies we surveyed, polymeric resins were the main class tested on real lignin-derived liquors, while zeolites, MOFs and COFs were mostly studied with single model compounds. Each class has strengths, but none has yet been shown to combine high capacity, selectivity, stability in the liquor and easy regeneration. Finding such a sorbent is the goal of our approach. We therefore propose a general computational sorbent screening framework. A short list of representative molecules from different feedstocks (for example softwood, hardwood and straw) and depolymerisation routes (oxidation or hydrogenation) is paired with candidate sorbents. Molecular simulations and quantum chemical calculations estimate how strongly each molecule binds to each sorbent, ranking the best combinations for experimental testing. The aim is a method that can be applied to any sorbent and any biorefinery stream.
Read moreView location in Venue Guide →Efficient Algorithms for Energy-Aware Single-Machine Scheduling with Battery Storage
TimeAuthorsAbstractEnergy-aware scheduling has become an increasingly important challenge in modern manufacturing environments, where coordinating production activities with electricity prices can substantially reduce operational costs and potentially contribute to lower carbon emissions. The integration of energy storage systems offers additional opportunities for cost savings by allowing electricity to be stored during low-price periods and discharged when prices are higher; however, doing so introduces additional combinatorial decisions, further complicating the already usually complex problem structure of scheduling problems. Despite its practical relevance, the algorithmic literature on energy-aware scheduling with energy storage remains relatively limited, with existing approaches predominantly relying on mathematical optimization models that can become computationally intractable for realistically sized instances. In this work, we study the problem of minimizing the energy cost of executing a set of jobs with sequence-dependent setup times on a single machine within a fixed time horizon, where electricity prices follow a Time-of-Use (TOU) tariff. In addition, we consider the presence of an industrial battery that can be charged from the grid and discharged during high-price periods to reduce overall energy costs. Even without energy storage, this scheduling problem is NP-hard; the introduction of a battery adds additional combinatorial interactions that further complicate the decision space. To tackle this problem, we develop two matheuristic algorithms that combine metaheuristic search with exact and heuristic optimization procedures to efficiently explore the solution space. In particular, the proposed approaches exploit different jobs scheduling and battery optimization strategies to balance computational efficiency and solution quality. Computational experiments demonstrate that one of the proposed algorithms consistently outperforms a benchmark MILP solver on medium- and large-sized instances, obtaining better solutions within substantially shorter computational times.
Read moreView location in Venue Guide →Measuring cardiac stiffness: integrated in vitro and in silico modeling of natural shear wave elastography
TimeAuthorsAbstractMyocardial stiffness is a key determinant of cardiac function and is altered in several cardiac pathologies. Natural cardiac shear wave elastography (SWE) offers a non-invasive approach to assess it by measuring the speed of shear waves generated by valve closure(1). However, interpreting SWE remains challenging because conventional shear wave speed (SWS)-stiffness relationships rely on assumptions that do not adequately describe the complex properties of the myocardium(2). We developed an integrated in vitro and in silico framework combining controlled SWE experiments with finite element simulations of natural shear wave propagation to investigate the influence of several confounders of shear wave propagation, such as excitation properties, material stiffness, geometry, and imaging view. In vitro, two left ventricular polyvinyl alcohol phantoms differing in stiffness were mechanically characterized (E = 40 and 69 kPa) and subjected to controlled local impacts using three different impactors. The resulting SWSs (2.68 and 2.99 m/s) and wave frequencies (~ 40 Hz) were consistent with reported in vivo ranges. In silico, the model calibrated using one impactor reproduced experimental SWS for two independent impactors, with absolute relative errors of 3.6–10.5%. Subsequent parametric analyses showed that increasing excitation duration from 5.9 to 23.6 ms reduced SWS from 2.91 to 2.32 m/s, while increasing wavelength from 65.6 to 76.6 mm, demonstrating that excitation characteristics affect the resulting wave field. Moreover, a 70% increase in phantom stiffness translated into a 35% increase in SWS, with experimental and simulated values remaining 32–41% below theoretical bulk-wave predictions derived from mechanical testing. These findings highlight the influence of excitation characteristics and geometrical confinement on natural shear wave propagation, challenging the direct application of conventional bulk-wave relationships to cardiac SWE. The proposed framework provides a basis for more accurate interpretation of myocardial stiffness measurements and future development of physiologically realistic cardiac SWE simulations. References 1 P. Santos et al., “Natural Shear Wave Imaging in the Human Heart: Normal Values, Feasibility, and Reproducibility,” IEEE Transactions on Ultrasonics, Ferroelectrics, and Frequency Control, 66, 3, 442–452, 2019. 2 A. Caenen et al., “Ultrasound shear wave elastography in cardiology,” JACC Cardiovasc. Imaging, 17, 3, 314–329, 2024.
Read moreView location in Venue Guide →Self-Supervised Learning for Sorting, Interpretation, and Quality Control of Microstructure Image Collections
TimeAuthorsAbstractHistorical materials-characterisation datasets often contain inconsistencies, incomplete metadata and imaging artefacts. As microscopy collections grow, manually organising and checking them becomes increasingly difficult. Within the AID4GREENEST project, we investigate self-supervised learning as a way to turn unlabelled scanning electron (SE) microscopy images of steel into structured resources for dataset exploration and quality assurance. We train image representations on 10,000 unlabelled steel microstructure images, avoiding the need for costly expert annotations. The resulting embedding space supports several tasks: similarity-based retrieval to find visually comparable microstructures across steel grades and processing routes; outlier screening to flag samples that may be corrupted, atypical or affected by preparation or imaging artefacts; and visual assessment of dataset coverage to help reveal imbalances and gaps. These outputs are intended to support expert review rather than replace it. We examine which image features influence the organisation of the learned space, helping assess whether the representations capture meaningful microstructural patterns. We also investigate how images from other characterisation modalities, including optical microscopy and electron backscatter diffraction (EBSD) image-quality maps, relate to a space trained primarily on SE images. The work covers ferritic, pearlitic, bainitic, martensitic and multiphase steels. By combining representation learning, retrieval and anomaly screening, this approach aims to make large microscopy collections easier to navigate and curate, while improving the traceability of quality-control decisions. The findings may inform imaging-based quality-assurance workflows beyond steel microscopy, particularly where large unlabelled datasets are available but annotated examples of anomalies are scarce.
Read moreView location in Venue Guide →