Periodic Reporting for period 1 - DIDEAROT (Digital Design strategies to certify and mAnufacture Robust cOmposite sTructures )
Periodo di rendicontazione: 2022-09-01 al 2024-02-29
New certified designs for structures are critical for the upcoming changes in conception of aircraft architectures. Breakthroughs in design and new strategies for a better use of material and integration of functions in aircraft are required. Digital conception and simulation will play a role to reach a certified design that includes production scenarios before full manufacturing.
The level of complexity of phenomena being solved through dedicated modelling techniques is constantly evolving and faces many challenges in validation and exploitation. For better use of these methods, scalability, speed and representativity need to be addressed. In the DIDEAROT project, ongoing until 2026, the design process is addressed using appropriate Machine Learning surrogates, benefiting from High Performance Computing.
The roadmap of the DIDEAROT project is at the crossroads for a holistic robust optimization of composite structures focused on digital predictions of two key aspects in its lifetime:
1) Manufacturing: predicting distortions, stress build-up and assembly challenges for ever-more integrated industrial scale composite parts
2) Dynamic loads and impact: predicting damage and effects from loads occurring at high speed or repeated loads over time that can lead to critical certification conditions.
Micro- and meso-models are computed and build upon relying on digital technologies (data or simulation driven) instead of using each one of these one-off for material simulation, high-quality models will be developed through the scales to ensure optimized design approaches up to an industrial scale.
For mechanical behaviour of composites including damage, this means that accurate representation of the evolution of the material properties are achieved at the ply and structural level using hybrid models trained through a database of microscale RVE models run on a multitude of load cases.
Manufacturing simulations use phenomenological curing models as building blocks. Hybrid models for the component level, capable of handling many design parameter evaluations will feed the full-scale simulations to reach robust design for manufacturing requirements. In parallel, improvements in the numerical schemes and high-performance computing of these multi-physical problems will finally validate the distortion and stress build-up in the structure.
The level of complexity of phenomena being solved through dedicated modelling techniques is constantly evolving and faces many challenges in validation and exploitation. For better use of these methods, scalability, speed and representativity need to be addressed. In the DIDEAROT project, ongoing until 2026, the design process is addressed using appropriate Machine Learning surrogates, benefiting from High Performance Computing.
The roadmap of the DIDEAROT project is at the crossroads for a holistic robust optimization of composite structures focused on digital predictions of two key aspects in its lifetime:
1) Manufacturing: predicting distortions, stress build-up and assembly challenges for ever-more integrated industrial scale composite parts
2) Dynamic loads and impact: predicting damage and effects from loads occurring at high speed or repeated loads over time that can lead to critical certification conditions.
Micro- and meso-models are computed and build upon relying on digital technologies (data or simulation driven) instead of using each one of these one-off for material simulation, high-quality models will be developed through the scales to ensure optimized design approaches up to an industrial scale.
For mechanical behaviour of composites including damage, this means that accurate representation of the evolution of the material properties are achieved at the ply and structural level using hybrid models trained through a database of microscale RVE models run on a multitude of load cases.
Manufacturing simulations use phenomenological curing models as building blocks. Hybrid models for the component level, capable of handling many design parameter evaluations will feed the full-scale simulations to reach robust design for manufacturing requirements. In parallel, improvements in the numerical schemes and high-performance computing of these multi-physical problems will finally validate the distortion and stress build-up in the structure.
The first 18 months of the DIDEAROT project focused on the foundations of methodologies while also bringing many interesting definitions of digital approaches & development of results.
Technical work has focused on reproducing state of the art results in new codes or with methodologies that will increase by several factors the speed at which complex evaluations will be made.
As all work packages have now started, leaders have taken on activities to specify the development framework:
- For work package 2 this includes the materials and the metrics of definition of the process induced distortion as well and the simulation of damage
- For work package 3 geometries, configurations, and applications of interest for the manufacturing simulations were defined as well as sensor technologies development. The first developments were tested on flat panel and corner bracket geometries
- For work package 4, damage models at different scales have been studied and developed both from a detailed simulation approach as well as machine learning approaches to evaluate material evolution more rapidly in stochastic loading scenarios
- For work package 5, a general review of the data and tests performed on vulnerability of composites was completed and a comprehensive definition of useful new information to input for digital models was reached.
Technical work has focused on reproducing state of the art results in new codes or with methodologies that will increase by several factors the speed at which complex evaluations will be made.
As all work packages have now started, leaders have taken on activities to specify the development framework:
- For work package 2 this includes the materials and the metrics of definition of the process induced distortion as well and the simulation of damage
- For work package 3 geometries, configurations, and applications of interest for the manufacturing simulations were defined as well as sensor technologies development. The first developments were tested on flat panel and corner bracket geometries
- For work package 4, damage models at different scales have been studied and developed both from a detailed simulation approach as well as machine learning approaches to evaluate material evolution more rapidly in stochastic loading scenarios
- For work package 5, a general review of the data and tests performed on vulnerability of composites was completed and a comprehensive definition of useful new information to input for digital models was reached.
In results for process simulation for manufacturing, progress past the state of the art include analysis of different numerical approaches to solve the problem of curing process for a composite part inside an autoclave. These numerical approaches will be classified in two distinctive groups based on transversal methodologies (AI & Machine learning and HPC). The first group will include hybrid-methods based on ML algorithms trained with synthetic/experimental datasets. While the second group will be focused on advanced simulations (direct FEMs) capable to perform a fully coupled fluido-thermo-structural simulation by making use of multiphysics HPC codes.
Simultaneously, two main progresses beyond SoTA are targeted for damage modelling:
• First, set-up accurate physics-based models capturing all damage mechanisms at micro and meso scales so to have a reliable source of synthetic data.
• Then, develop material specific AI/ML methodologies, to be used in conjunction with the hybrid methods of Trans 1 theme, allowing to speed-up multi-scale modelling by 5 order of magnitude while keeping sufficient accuracy.
For AI & Machine Learning, the consortium is analysing strategies to reduce the complexity of the classical approximations for solving parametric (differential) physical equations by means of FEMs.
Conjugation of reduced order models (ROM not only limited to classical linear methods but including non-linear ones based on machine inference) with FEM models, study of temporal relationships in simpler vector spaces, down-sampling techniques or metamodels are being used.
On the front of HPC, DIDEAROT progresses beyond the state of the art exploring different numerical methodologies for the design cycle of aircraft structures taking benefit of existing frameworks based on HPC.
Last but not least, the use of HPC will not only be applied to solve large-scale problems, but also to generate the synthetic dataset for training the machine learning models (hybrid methods) for both process & damage modelling. These approaches will contribute to the generation of data for further advanced models based on machine learning.
Simultaneously, two main progresses beyond SoTA are targeted for damage modelling:
• First, set-up accurate physics-based models capturing all damage mechanisms at micro and meso scales so to have a reliable source of synthetic data.
• Then, develop material specific AI/ML methodologies, to be used in conjunction with the hybrid methods of Trans 1 theme, allowing to speed-up multi-scale modelling by 5 order of magnitude while keeping sufficient accuracy.
For AI & Machine Learning, the consortium is analysing strategies to reduce the complexity of the classical approximations for solving parametric (differential) physical equations by means of FEMs.
Conjugation of reduced order models (ROM not only limited to classical linear methods but including non-linear ones based on machine inference) with FEM models, study of temporal relationships in simpler vector spaces, down-sampling techniques or metamodels are being used.
On the front of HPC, DIDEAROT progresses beyond the state of the art exploring different numerical methodologies for the design cycle of aircraft structures taking benefit of existing frameworks based on HPC.
Last but not least, the use of HPC will not only be applied to solve large-scale problems, but also to generate the synthetic dataset for training the machine learning models (hybrid methods) for both process & damage modelling. These approaches will contribute to the generation of data for further advanced models based on machine learning.