01
Design-space learning and Parametric Model Embedding
High-dimensional parameterizations allow engineering geometries to be described with considerable flexibility, but they also increase the cost and difficulty of design exploration and optimization. This research studies compact representations of parametric design spaces that retain the variability required by the engineering problem.
Andrea Serani is the lead developer of Parametric Model Embedding (PME). PME constructs a reduced representation while preserving an explicit mapping between reduced coordinates and the original design variables, allowing the established parameterization to remain part of the design workflow.
02
Physics-aware and nonlinear representations
The PME research line extends dimensionality reduction beyond geometric information alone. Physics-informed formulations incorporate physical observables into the embedding, while physics-driven developments use physical evidence to guide the representation of the design space.
Current work investigates nonlinear extensions of PME and design-manifold learning. The objective is to represent nonlinear structure in complex design spaces while maintaining a usable relationship with the engineering parameterization and the quantities needed for analysis and optimization.
03
Simulation-based, surrogate and multi-fidelity optimization
Simulation-based design optimization often relies on computationally expensive analyses. This research develops surrogate modelling, adaptive sampling, and multi-fidelity strategies that combine information from simulations with different costs and accuracies.
The programme also addresses uncertainty in operating conditions and computational outputs. These methods support design decisions in problems where direct high-fidelity exploration is impractical, particularly in multidisciplinary vehicle design.
04
Scientific machine learning and reduced-order approaches
Scientific machine learning is used to integrate geometry, physical information, simulation data, and predictive models. The work includes reduced representations for design exploration, data-driven models for engineering analysis, and methods that connect learning with optimization rather than treating prediction as an isolated task.
Related research includes reduced-order and time-series approaches for computational engineering problems, with an emphasis on interpretable representations and their role in design decisions.
05
Engineering applications
The methods are studied through applications in marine, naval, underwater, and aerospace vehicle design. These problems combine computational fluid dynamics, geometric parameterization, multiple physical disciplines, uncertain conditions, and competing design objectives.
Collaborative projects, group software, datasets, and opportunities are maintained by the MAO Research Group.
Selected references
- 2026Physics-informed dimensionality reduction for propeller shape optimization
- 2025Extending Parametric Model Embedding with Physical Information for Design-Space Dimensionality Reduction in Shape Optimization
- 2026A Survey on Design-Space Dimensionality Reduction Methods for Shape Optimization
- 2023Parametric model embedding