The student Julen Mendicute San Martin obtained an OUTSTANDING qualification with 'CUM LAUDE' mention

Back

The student Julen Mendicute San Martin obtained an OUTSTANDING qualification with 'CUM LAUDE' mention

THESIS

The student Julen Mendicute San Martin obtained an OUTSTANDING qualification with 'CUM LAUDE' mention

2022·07·12

$titulo.getData()


Thesis title: "Predicción del comportamiento a impacto de materiales compuestos fabricados mediante Resin Transfer Moulding aplicando Gemelos Digitales basados en Machine Learning"

Court:

  • Chairmanship: Carlos Daniel González Martínez (Universidad Politécnica de Madrid)
  • Vocal: Isabel Harismendy Ramírez de Arellano (Tecnalia Research and Innovation)
  • Vocal: Juan Antonio García Manrique (Universitat Politècnica de València)
  • Vocal: Germán Castillo López (Universidad de Málaga)
  • Secretary: Laurentzi Aretxabaleta Ramos (Mondragon Unibertsitatea)

Abstract:

Carbon fibre reinforced polymers (CFRP) have proven to be effective for structural lightweighting as they combine low density with good mechanical properties, especially in terms of specific strength and stiffness. Resin Transfer Moulding (RTM) is an effective process in the manufacture of high performance, geometrically complex structural parts with low process costs. However, the robustness of the RTM process remains a challenge, as both material and process uncertainties negatively affect the impregnation quality, generating void defects and dry zones. To address these defects, a paradigm shift has to take place, as the production system itself has to become an "expert" in material science and process technology. The manufacturing system must consider the defects generated in impregnation stage to recognise the risks in the final structural performance. In other words, a Digital Twin (DT) of the physical process must be constructed and accurately recreate the RTM process. For this purpose, the continuous Process-Structure-Property-Performance (PSPP) modelling approach has proven to be the key to the generation of DTs. Therefore, the objective of this thesis is to generate the scientific-technological knowledge for Digital Twins of the RTM process following the PSPP approach.

Firstly, an experimental characterisation of the material based on the PSPP approach has been carried out to quantify the effect of process parameters on void generation (Process-Structure), and subsequently, void effect on impact properties (Structure-Property). For this purpose, NCF-Epoxy biaxial laminates (0/90) with different void contents (0.58%, 1.35%, 2.44% and 4.34%) have been tested by the drop-weight impact test method. Samples with high void content (4.34%) recorded a 25.88% reduction in peak force, and a 9.57% reduction in dissipated energy compared to samples with low void content (0.58%).

Secondly, a Machine Learning (ML) based impregnation quality diagnosis model has been developed. This model has been fed with information from process and has been able to predict the generated impregnation structure (Process-Structure). Among the different predictive models studied, Extreme Gradient Boosting and Light Gradient Boosting Machine were the most accurate models for predicting RTM filling quality, with an accuracy of 84.9% and 83.35%, respectively. In addition, a scaling of the ML model has allowed not only to predict the part quality, but also to locate the zones where the defect was generated. Comparing the computational time of the ML model with the FEM model, a reduction in computational time from 360 s to 1 s was observed. Thus, supervised learning predictive models are fast enough to be integrated into the process and feed the digital twin of the process.

Finally, a surrogate model based on ML-FEM has been developed, which, knowing the location and position of the impregnation effect, is able to predict the reduction of mechanical properties (Structure-Properties), and then the impact performance of the parts (Properties-Performance). The surrogate model has allowed predictions for (i) quality classification based on the maximum displacement, (ii) numerical prediction (regression) of the maximum displacement, and (iii) multi-output regression to obtain the impact curves (F(t), E(t) and U(t)). The results obtained have shown an accuracy of 95.85% in the classification and a R2 of 0.86 in the regression. Once the surrogated model is trained, it is able to make predictions in less than 5 seconds, unlike the FEM model that needs about 60 minutes.

Combining both the impregnation quality diagnostic model (Process-Structure) and the structural performance predictive model (Structure-Properties-Performance), it has been possible to generate a DT based on the PSPP approach. Both models are fast enough to be integrated into the RTM process and allow online structural validation considering the process defects.