Abstract
MORDICUS is the French acronym for “Méthodes d’Ordre RéDuIt Conçues pour des Usages induStriels”, translated to “reduced-order methods designed for industrial usage”. It is the name of a collaborative project that took place from 2018 to 2023, with the objective of developing a standard for a datamodel and basic computational treatment for reduced-order modeling in the French community.
Introduction
High-Fidelity Models (HFM), such as large-scale Finite Element Method (FEM) simulations, are the backbone of modern engineering but are notoriously slow and computationally expensive. Reduced-Order Modeling (ROM) offers a way to create "surrogate" models that run orders of magnitude faster while maintaining accuracy. However, two major hurdles have slowed industrial adoption: a lack of standardization across different data structures and the "intrusivity" of most ROM techniques, which often require modifying the source code of underlying solvers.
The MORDICUS project, a French consortium including EDF and Safran, addresses the standardization issue, while the GenericROM library provides a nonintrusive path for industrial applications, allowing engineers to utilize proprietary commercial solvers like Abaqus or Z-set without internal code modification.
Methods and Framework Architecture
The framework relies on several key pillars to ensure interoperability and efficiency:
The MORDICUS Datamodel
A standardized Python/C++ structure organized into three main classes:
- CollectionProblemData: The meta-structure managing the overall project data.
- ProblemData: Defines the physics, geometry, and loading conditions.
- Solution: Stores the snapshots and reduced coordinates.
Computational Techniques
- Snapshot POD (Proper Orthogonal Decomposition): The primary method for data compression, identifying the most significant "modes" of a system to reduce the number of degrees of freedom.
- ECM (Empirical Cubature Method): A hyper-reduction technique used to speed up the assembly of nonlinear operators by evaluating the physics only at a small subset of integration points.
- Nonintrusive Workflow: GenericROM utilizes an external finite element engine (
BasicTools) to handle operator assembly. This allows it to build ROMs from snapshots generated by commercial solvers without needing access to the solver's internal source code.
Industrial Application: The "Cube" Case
The framework's capabilities are demonstrated using a representative industrial test case: a 1m-wide cube subjected to complex, time-dependent loading. This includes variable thermal fields, centrifugal effects (rotation), and pressure. The cube employs an elastoviscoplastic material model with nonlinear hardening and Norton flow, reflecting the high complexity found in turbine blade simulations. High-fidelity snapshots were generated using the Z-set solver and imported into the MORDICUS format for reduction.
Key Insights
- Standardization as an Enabler: By defining a common datamodel, the MORDICUS project allows industrial partners using different solvers to share ROM algorithms and workflows seamlessly.
- The Power of Nonintrusivity: The ability to use pre-computed databases of snapshots from commercial software bypasses the need for expensive custom implementations within proprietary solvers.
- Precomputation for Efficiency: Real-time speedup is achieved in the offline stage by precomputing quantities like the strain tensor applied to the Reduced-Order Basis (ROB), reducing the online stage to simple, dense matrix operations.
Results and Performance
In structural mechanics tutorials, the framework maintained high fidelity, with the relative projection error of the displacement field kept below $1 \times 10^{-5}$. The framework successfully transforms large, sparse linear systems into small, dense systems, significantly reducing the computational burden. Qualitative comparisons show that the ROM's predicted displacement and accumulated plasticity fields are virtually indistinguishable from high-fidelity references.
Limitations and Future Work
While the framework is nonintrusive, certain tutorials still require licenses for specific constitutive law solvers. Currently, the library is optimized for nonlinear quasistatic structural mechanics and transient thermal problems. Future expansions may target fluid dynamics or electromagnetics. Additionally, while the framework can handle nonparametrized variability, this introduces a computational trade-off that remains a focus for future optimization.
Summary updated