DCs & Workpackages
DC1 will work on the development of efficient techniques for updating the knowledge on the latent parameters of a physics-based model that is used for predicting the performance of critical components. In particular, they will focus on the challenges posed by the limited and mixed type of information arising from the manufacturing process and from limited experimental observations on the critical component once produced. In particular, DC1 will investigate the extension of the Bayesian Inference framework to deal with deep uncertainties in the information required to build the prior distribution and of the likelihood function. This will result in an innovative approach combining physics-based models and statistical machine learning techniques to yield an accurate description of the so-called posterior distribution of the uncertain input parameters. In particular, this project will require the development of a generalised description of the prior information and of the likelihood distribution. In turn, this will pose an additional challenge on how to evaluate efficiently the posterior distribution. DC1 will therefore focus on the development of advanced Monte Carlo-based sampling techniques.
DC2 will work on the development of efficient techniques for assessing the effects of uncertainties in the input parameters of physics-based models on the design performance. In particular, DC2 will investigate the combination of probabilistic and non-probabilistic approaches (such as interval, convex and fuzzy descriptions) to describe the different levels of information available at the design stage on the uncertain parameters (e.g. only the bounds a variable might take, a probability density function with imprecisely known parameters). They key challenge here is that the quantification of the resulting uncertainty in the performance metric might require a large number of simulations of expensive-to-run physics-based models. Surrogate models might be used to approximate the relationship between the response and the input parameters to yield a model that is computationally cheaper to run. However, these approaches would require evaluating a large number of runs to construct an accurate surrogate model, and to fine-tune the surrogate model parameters. Therefore, they can entail a considerable computational effort if a large number of uncertain parameters need to be considered. DC2 will work on the development of efficient uncertainty propagation strategies aimed at (i) reducing the number of the expensive-to-evaluate model runs required to yield accurate estimates of the performance under uncertainties, and (ii) quantifying the errors on these estimates that would arise when considering a limited number of simulations.
Future seafastening techniques within the offshore industry will rely more on friction-based connections since transported objects are exceeding vessel sizes. These connections merely rely on contact mechanisms between the transported structure and the support elements such as friction pads or rollers. However, the lack of high-fidelity guidelines for the design, characterization and selection of these support elements lead to higher costs, reduced safety and limitations for the transport and installation operations. To fill this gap, DC3 will develop testing and predictive models to quantify the variability of the interface features, such as the friction force and the interfacial stiffness, for friction-based support elements. The input parameters for this model will be the surface roughness, the material property of the friction-based element, interfacial medium (water, salt), ambient medium (temperature) and the environmental and operational loads. The developed model will have a dual function to improve the performance prediction of the support element and the required manufacture tolerances in terms of roughness and material properties based on calculated representative environmental and operational loads, and to serve as a reference model to generate training data for the development of a grey-box model for the stability of the seafastened object. Besides the modelling part, the DC will carry out an extensive experimental campaign to characterize the interface behaviour of such friction pads. The virtual training data set obtained by the numerical model will be enhanced and complemented by means of the experimental campaign.
DC4 will develop strategies to quantify the effects of local variations in density and geometry compared to the nominal CAD design on the stop-band behaviour of 3D-printed locally resonant metamaterials. The spread on the designed stopband location and width will be assessed for different metamaterial designs, focusing on acoustic as well as vibrational resonances. Uncertainties on geometry, density and stiffness distributions will be quantified. Design rules will be derived for different resonator geometries, linking geometrical dimensions in the CAD file via the production process to a quantified spread on the expected result, accounting for the existing uncertainties. Bayesian approaches will be used to this extent. Having an established link between the nominal CAD design and performance, while accounting for manufacturing uncertainties will provide a leap forward to further industrial applications of metamaterials.
ESR5 will investigate artificial sound sources designed to warn pedestrians of their approach. The level of the noise for such sources in the forward direction is determined by the requirements of audibility. The drive-by noise, however, is also determined by the directivity of the source of warning sounds. This project would consider designs in which the directivity was optimised to minimise the drive-by noise, while maintaining an appropriate level in the forward direction. In particular, a novel directional line-implemented sound source will be considered.
Innovative aspects: novel sound emitting device development with enhanced directivity
In additive manufacturing the definition of the manufacturing strategy is difficult, and typically for a complex part a few iterations are necessary, involving actual manufacturing of the part, quality control and adaptation of the process input parameters. This is costly and time consuming, however, necessary for critical components. Consequently, manufacturing and performance simulations, and the use big data during the design and build preparation phase are very valuable. The objectives of DC6 are:
- Propose a validation strategy for the manufacturing simulation models for additive manufacturing.
- Validate the manufacturing simulation (physical and data-driven) models that simulate the actual building process through the use of online monitoring tools such as highspeed thermal cameras.
- The output of the simulation models needs to be interpreted in order to take meaningful decisions on the process input parameters and therefore the DC will propose algorithms and tools for interpretation of the manufacturing simulation outcome.
- Establish and validate a workflow that makes it possible to interpret a manufacturing simulation and take corrective actions during the preprocessing phase (e.g., manufacturing simulation outputs thermal stresses, corrective action is proper positioning of support structures in metal AM).
- Apply the work-flow as a proof-of-principle on a minimum of 3 industrial cases.
The overall task is to assess how the manufacturing-induced residual stresses, plastic strains and geometry imperfections influence vibration performance of thin-walled spatially curved panels (possibly, under heavy fluid loading), used as components of large centrifugal pumps. In an assembled structure, these components act both as resonators, which generate tonal noise at excessive levels, and as waveguides, which transmit the vibro-acoustic energy from the source (rotating fluid-loaded blades) to the outer parts of the piping system. It is anticipated that a high fidelity modelling of sheet metal forming by means of advanced numerical tools will provide the necessary information to identify the eigenfrequencies and eigenmodes of these components and to predict distribution of structural intensities at the dominant excitation frequencies. In the course of modelling, sensitivities of dynamic characteristics of panels to parameters of the manufacturing process will be assessed. The intended work will also encompass the relative uncertainty assessment of the eigenfrequencies, eigenmodes and structural intensities in view of performance of assembled large centrifugal pumps under uncertain boundary conditions.
DC8 will develop models that couples manufacturing and performance simulations into a unified model in which consequences of changes of meaningful parameters can be immediately predicted (digital twin). The main input parameters are related to the manufacturing and operation processes, while the output is centred around the fatigue performance.
First, DC8 will develop and validate performance and manufacturing models for various inputs individually and second, combine the models to a one-way coupled manufacturing-performance simulation such that material hardening, geometry, residual stresses, plastic strains and tolerances in the manufacturing may be accounted for in the performance simulations. Last, the model is extended to a fully coupled model such that design optimisation accounts for the manufacturing processes in the evaluation of the product performance. There will be a specific focus on simulation-based reduced-order models as well as hybrids (simulation and data) to allow also for real-time calculations.
DC9 will develop active learning approaches and strategies to enhance explainability of learned models and provide insights on the classification models' prediction to reduce the manual labelling effort. The effectiveness of such approaches and strategies will be evaluated on real-world use cases from the manufacturing industry. The use of active learning for automated visual inspection of manufactured products will advance current defect detection, by eliminating the manual inspection for most of the produced pieces, while requiring a labelling effort only on those that are most informative to the model. In addition, the use of explainable artificial intelligence will provide means towards understanding the rationale behind the models' decisions, which can be used in two ways: (a) by the machine learning engineers to enhance future models, and (b) by the operators, to obtain hints regarding why a piece could be considered defective. We envision that the implementation of such an approach will speed up defect detection, reduce manual labour, and enrich the operators experience with additional insights that are currently not available to them.
DC10 will develop active learning approaches in manufacturing processes, as a means of facilitate human-robot (human-machine) interactions and accelerating knowledge acquisition. A joint knowledge base for machines and humans will be created and will be continually updated based on active learning approaches. The potential of updating the knowledge base with results based on simulated reality interactions and with knowledge derived from formalizations (e. g., the ontological descriptions) of explainable AI models will be studied and implemented. The knowledge base will facilitate faster and more accurate decisions (e. g., in the process of troubleshooting or reparation on a shop floor).
Active learning approaches will be tailored towards the needs of the manufacturing domain. The work will involve customization and enhancements to the ontologies of the systems, as well as to the development of a semantic query engine that will enable the interaction between humans and machines. The system should enable the machine to query human experts about quality management and agile manufacturing, with the usage of relevant interfaces and interaction mechanisms such as NLP (Natural Language Processing). The work will research the potential of the implementation of active learning at all levels of a production process and of manufacturing enterprises from devices on the shop floor to the global supply chains. Knowledge formalization will be based on existing formalized knowledge descriptions (e. g., general and specific ontologies translated in knowledge graphs).