Periodic Reporting for period 1 - PorMatDesign (Machine learning-aided multiscale design of porous materials tailored to application-specific, hydro-mechanical performance requirements)
Período documentado: 2023-10-01 hasta 2025-09-30
PorMatDesign set out to develop a computationally feasible framework for the multi-scale design and optimization of porous materials, whose microscopic structure determines their overall mechanical and hydraulic performance. Such materials are central to technologies ranging from energy storage and filtration to additive manufacturing and biomedical engineering. However, designing them remains a major scientific challenge: the relationships between their pore-level geometry and their large-scale properties are nonlinear, multi-physical, and span several spatial scales. This makes direct optimization computationally prohibitive.
The project addressed this problem by integrating machine learning and physics-based simulation into a unified optimization framework. Rather than relying solely on trial-and-error or high-fidelity numerical models, PorMatDesign explored how physics-informed neural networks (PINNs) and advanced optimization strategies could be used to link structure, process, and performance more efficiently.
The central objective was to create a tractable simulation–optimization pipeline that can generate porous microstructures tailored to specific target behaviors—such as flow permeability, stiffness, or energy absorption—while maintaining physical realism. This required:
- developing mathematical models for reducing the dimensionality of porous-structure representations,
- designing multi-scale neural networks that combine information across pore-scale and Darcy-scale physics,
- and formulating a topology optimization algorithm capable of searching large design spaces efficiently.
The project’s methodological advances were implemented in an open-source software tool called 'Poromotiv', which integrates porous-media generation, descriptor computation, and optimization within a single, reproducible workflow.
The project addressed this problem by integrating machine learning and physics-based simulation into a unified optimization framework. Rather than relying solely on trial-and-error or high-fidelity numerical models, PorMatDesign explored how physics-informed neural networks (PINNs) and advanced optimization strategies could be used to link structure, process, and performance more efficiently.
The central objective was to create a tractable simulation–optimization pipeline that can generate porous microstructures tailored to specific target behaviors—such as flow permeability, stiffness, or energy absorption—while maintaining physical realism. This required:
- developing mathematical models for reducing the dimensionality of porous-structure representations,
- designing multi-scale neural networks that combine information across pore-scale and Darcy-scale physics,
- and formulating a topology optimization algorithm capable of searching large design spaces efficiently.
The project’s methodological advances were implemented in an open-source software tool called 'Poromotiv', which integrates porous-media generation, descriptor computation, and optimization within a single, reproducible workflow.
PorMatDesign was carried out at the University of Luxembourg between October 2023 and September 2025, with a three-month secondment at the University of Strasbourg. The work progressed through three complementary stages:
1. Methodological consolidation: A comprehensive review of existing methods for embedding physics into neural networks and for hybridising classical and quantum machine learning was conducted to identify computationally feasible routes for multi-scale porous-media optimisation. This work resulted in two major review papers currently under peer review.
2. Tool development: The project produced a new open-source Python library called Poromotiv, which provides a unified environment for generating, analysing, and optimising 3D porous microstructures. Poromotiv integrates several types of porous-media models—statistical (LCRF), geometric (Boolean, Voronoi), and image-based (micro-CT)—within a common data representation. The library automates descriptor calculation, physical property estimation, and topology optimisation, ensuring full reproducibility through standardized metadata and file structures.
3. Multi-scale modelling and optimisation: The project demonstrated how physics-informed neural networks (PINNs) can couple pore-scale and Darcy-scale models of fluid flow, allowing information to propagate across scales while maintaining physical continuity. Building on this framework, a reduced-parameter topology-optimisation algorithm was implemented to efficiently search the morphological design space for desired hydro-mechanical performance targets.
The combination of theoretical analysis, algorithmic development, and open-source implementation makes PorMatDesign a reference point for future research on the physics-informed design of complex materials.
1. Methodological consolidation: A comprehensive review of existing methods for embedding physics into neural networks and for hybridising classical and quantum machine learning was conducted to identify computationally feasible routes for multi-scale porous-media optimisation. This work resulted in two major review papers currently under peer review.
2. Tool development: The project produced a new open-source Python library called Poromotiv, which provides a unified environment for generating, analysing, and optimising 3D porous microstructures. Poromotiv integrates several types of porous-media models—statistical (LCRF), geometric (Boolean, Voronoi), and image-based (micro-CT)—within a common data representation. The library automates descriptor calculation, physical property estimation, and topology optimisation, ensuring full reproducibility through standardized metadata and file structures.
3. Multi-scale modelling and optimisation: The project demonstrated how physics-informed neural networks (PINNs) can couple pore-scale and Darcy-scale models of fluid flow, allowing information to propagate across scales while maintaining physical continuity. Building on this framework, a reduced-parameter topology-optimisation algorithm was implemented to efficiently search the morphological design space for desired hydro-mechanical performance targets.
The combination of theoretical analysis, algorithmic development, and open-source implementation makes PorMatDesign a reference point for future research on the physics-informed design of complex materials.
PorMatDesign advanced the state of the art in three significant ways:
- Unified generative and optimisation framework: For the first time, several previously separate porous-media model families have been integrated into a single, descriptor-based computational architecture, allowing direct comparison and benchmarking across stochastic, geometric, and image-based methods.
- Physics-informed multi-scale learning: The project established a practical formulation for coupling scale-specific PINNs through physics-based interface losses, reducing the computational burden of multi-scale simulation without sacrificing physical fidelity.
- Open, reproducible research infrastructure: All workflows developed in the project are designed for transparency and reuse, providing the research community with a sustainable platform for extending, validating, or repurposing the methods.
The results contribute to Europe’s scientific and technological leadership in AI-driven materials design, supporting the digital transformation of the materials sector. Potential beneficiaries include research groups, engineering consultancies, and industries working on additive manufacturing, filtration, and energy systems. In the long term, the methods developed can reduce the environmental and financial costs of trial-and-error material development, aligning with the goals of the European Green Deal and the Digital Europe strategy.
- Unified generative and optimisation framework: For the first time, several previously separate porous-media model families have been integrated into a single, descriptor-based computational architecture, allowing direct comparison and benchmarking across stochastic, geometric, and image-based methods.
- Physics-informed multi-scale learning: The project established a practical formulation for coupling scale-specific PINNs through physics-based interface losses, reducing the computational burden of multi-scale simulation without sacrificing physical fidelity.
- Open, reproducible research infrastructure: All workflows developed in the project are designed for transparency and reuse, providing the research community with a sustainable platform for extending, validating, or repurposing the methods.
The results contribute to Europe’s scientific and technological leadership in AI-driven materials design, supporting the digital transformation of the materials sector. Potential beneficiaries include research groups, engineering consultancies, and industries working on additive manufacturing, filtration, and energy systems. In the long term, the methods developed can reduce the environmental and financial costs of trial-and-error material development, aligning with the goals of the European Green Deal and the Digital Europe strategy.