In aerospace engineering, the simulation of propulsion systems—whether for gas turbine engines, rocket nozzles, or hypersonic scramjets—remains a critical pillar of design and certification. High-fidelity models, such as large-eddy simulations (LES) or direct numerical simulations (DNS), resolve fine-scale physics like turbulence and combustion chemistry with remarkable accuracy. However, the computational cost of these approaches can be staggering: a single combustor simulation can consume tens of thousands of core-hours on a high-performance computing cluster. This bottleneck severely limits the number of design iterations that engineers can perform, slowing innovation and driving up development costs. Reduced-order modeling (ROM) offers a pragmatic path forward. By compressing the essential physics of a system into a low-dimensional representation, ROM allows engineers to run simulations in seconds or minutes rather than days, while preserving the fidelity needed for design decisions. This article explores how ROM is reshaping propulsion system simulations, from the core mathematical techniques to practical applications and emerging trends.

Understanding Reduced-Order Modeling

Reduced-order modeling refers to a class of techniques that produce simplified, low-dimensional approximations of high-fidelity dynamical systems. The underlying idea is that many complex physical phenomena exhibit low-rank structure—that is, their behavior can be accurately captured using a relatively small number of dominant modes or patterns. ROM identifies these modes from either physics-based reasoning (e.g., modal analysis) or data-driven methods (e.g., machine learning), then constructs a much smaller system of equations that approximates the original model’s input-output behavior.

At its core, a ROM replaces a high-dimensional state vector (often with millions of unknowns) with a compressed representation of, say, 10 to 100 degrees of freedom. This compression is achieved through a projection onto a reduced subspace. The key steps are:

  1. Snapshot generation: Run the full-order model (FOM) at a set of representative operating conditions and collect snapshots of the state variables (e.g., pressure, temperature, velocity fields).
  2. Basis construction: Decompose the snapshot matrix to extract a set of basis functions that capture most of the energy or variance.
  3. Projection: Project the governing equations (e.g., Navier-Stokes, energy conservation) onto the reduced subspace to obtain a low-dimensional ODE or algebraic system.
  4. Time integration or evaluation: Solve the reduced system for new inputs or parameters orders of magnitude faster than the FOM.

The result is a model that retains the essential physics—vortex shedding, shock-wave interactions, combustion heat release—while discarding unnecessary degrees of freedom. This enables rapid what-if analyses, parametric studies, and real-time control.

Key Techniques in Reduced-Order Modeling for Propulsion

Several ROM methodologies have proven especially effective in propulsion applications. Each technique trades off accuracy, computational cost, and interpretability.

Proper Orthogonal Decomposition (POD)

Proper Orthogonal Decomposition, also known as Karhunen-Loève expansion, is the most widely used ROM approach in fluid dynamics. POD identifies spatial modes that optimally represent the energy content of the flow. The modes are obtained from a singular value decomposition (SVD) of the snapshot matrix. For propulsion simulations, POD has been applied to compress turbulent flow fields in combustors and turbine stages. A POD-ROM projects the full-order Navier-Stokes equations onto a subspace spanned by the dominant POD modes, resulting in a system of ordinary differential equations (ODEs) that can be integrated with standard solvers. The limitation is that POD-ROMs are typically linear and may struggle with strongly nonlinear phenomena or parameter variations not included in the training snapshots.

Dynamic Mode Decomposition (DMD)

Dynamic Mode Decomposition is a data-driven technique that extracts coherent spatiotemporal structures—each associated with a distinct frequency and growth rate. Unlike POD, which ranks modes by energy, DMD extracts modes based on their dynamic behavior. This makes DMD ideal for capturing oscillatory phenomena such as combustion instabilities, flame-acoustic coupling, or rotor-stator interactions in turbomachinery. DMD-ROMs can predict the evolution of the system over time without requiring explicit governing equations, making them valuable when the underlying physics is not fully known or too expensive to solve numerically. Researchers at institutions like NASA Glenn Research Center have used DMD to build reduced-order models for rocket injector dynamics and to identify modes that drive high-frequency combustion oscillations.

Autoencoder-Based Neural Networks

Recent advances in machine learning have introduced autoencoders—neural networks that learn a compressed latent representation of high-dimensional data. An autoencoder consists of an encoder that maps the full state to a low-dimensional latent vector and a decoder that reconstructs the original state. By training on a diverse set of simulation snapshots, the autoencoder learns a nonlinear manifold that can represent the system more efficiently than linear techniques like POD. For propulsion simulation, autoencoder ROMs have been used to model spray combustion and turbulent mixing, capturing sharp gradients and nonlinear interactions that linear ROMs miss. With modern GPU hardware, autoencoder-based ROMs can run near real-time, enabling applications in digital twins and online monitoring.

Benefits of ROM for Propulsion System Development

The adoption of reduced-order modeling in propulsion engineering delivers several concrete advantages, particularly in the context of modern, time-to-market-driven development cycles.

  • Drastically Reduced Simulation Time: ROMs typically reduce computation time by factors of 100 to 10,000. A LES that takes a week on a 1,000-core cluster can be approximated by a ROM that runs in under a minute on a single workstation. This acceleration allows engineers to explore design spaces that were previously intractable.
  • Cost Efficiency: By minimizing reliance on expensive high-performance computing resources, ROMs lower the overall cost of simulation-driven design. Smaller companies and research groups gain access to high-quality predictive capabilities without needing supercomputing centers.
  • Design Optimization: ROMs enable parametric sweeps and gradient-based optimization over multiple variables (e.g., nozzle geometry, fuel injection timing, blade twist). A full-order optimization might require only a handful of evaluations; with ROM, thousands of evaluations become feasible, leading to more thoroughly optimized designs.
  • Real-Time and Embedded Applications: Reduced-order models can be deployed in hardware-in-the-loop testing, engine control units, or flight simulators. For example, a ROM of a gas turbine combustor can provide real-time predictions of temperature profiles and emissions, enabling active control strategies to avoid lean blowout or reduce NOx formation.
  • Enhanced Physical Insight: The low-dimensional nature of ROMs often reveals which physical mechanisms dominate system behavior. The dominant POD modes, for instance, may show the primary heat release zones or the most responsive flow structures, guiding engineers toward targeted design changes.

Applications Across Propulsion Disciplines

Reduced-order modeling has found traction in nearly every subfield of propulsion engineering, from conventional gas turbines to advanced rocket systems and emerging electric propulsion concepts.

Gas Turbine Engines

In turbofan and turbojet engines, ROMs are applied to:

  • Combustor design: ROMs approximate the complex turbulent reacting flow within a combustor to predict flame stability, pollutant formation (NOx, CO), and liner temperature distribution. This enables rapid iteration over fuel injector geometries and dilution hole patterns.
  • Turbine blade cooling: High-fidelity conjugate heat transfer simulations of internally cooled blades are replaced by ROMs that capture the coupling between hot gas path, film cooling, and internal channel flow, accelerating the design of more efficient cooling schemes.
  • Compression system stability: ROMs derived from full-annulus CFD predict surge and stall margins for multistage compressors, allowing engineers to optimize variable guide vanes and bleed valves without exhaustive full-order campaigns.

Rocket Propulsion

Liquid and solid rocket engines benefit from ROM for:

  • Combustion instabilities: DMD-based ROMs identify the acoustic modes that drive pressure oscillations in thrust chambers. By linking these modes to injector characteristics, engineers can design damping devices or shift resonance frequencies before costly hot-fire tests.
  • Nozzle flow dynamics: ROMs capture the shock-boundary layer interactions inside supersonic nozzles, predicting thrust loss due to flow separation. During the development of the SpaceX Raptor engine, ROM-like approaches have been used to rapidly evaluate nozzle extension designs under varying chamber pressures.
  • Plume simulations: Simplified plume models based on ROM provide fast predictions of base heating and plume impingement on launch vehicle structures, supporting trade studies for stage separation and landing configurations.

Electric and Hybrid Propulsion

Emerging electric propulsion systems (e.g., Hall thrusters, ion engines) also benefit from ROM. For example, a reduced-order model of the plasma discharge in a Hall thruster can predict thrust and specific impulse across a range of voltage and propellant flow conditions, replacing particle-in-cell simulations that can take weeks to run. These ROMs facilitate the design of power processing units and thruster control algorithms for satellite station-keeping and deep-space missions.

Overcoming Critical Challenges

Despite its promise, ROM has not been universally adopted across the propulsion industry, largely due to several persistent challenges that active research aims to address.

Accuracy across varying operating conditions: A ROM built using snapshots from a narrow range of conditions (e.g., one fuel-air ratio and inlet temperature) may perform poorly when extrapolated to different regimes. This limits its use for full flight envelope coverage. Multi-fidelity ROMs that combine data from a few high-fidelity simulations with many lower-fidelity runs (e.g., using CFD of varying mesh resolution) are one promising approach. Another is the use of Galerkin projection with adaptive basis methods that update the reduced basis on the fly when the system deviates from trained conditions.

Handling strong nonlinearities: Combustion involves highly nonlinear phenomena—turbulent mixing, chemical kinetics, heat transfer. Linear projection-based ROMs (e.g., POD-Galerkin) often fail to capture these dynamics. Solutions include coupling the ROM with a local, physics-based subgrid model (such as a flamelet model) or using nonlinear manifold learning via autoencoders and operator inference. Techniques like neural ordinary differential equations (neural ODEs) are being explored to learn the reduced dynamics directly from data.

Integration with multi-physics simulations: Propulsion systems couple fluid dynamics, structural mechanics, thermal response, and even electromagnetics (in electric propulsion). Creating a global ROM that captures all interactions is a formidable task. Domain decomposition approaches—building separate ROMs for each physical domain and coupling them through shared interfaces—have shown promise. Alternatively, proper generalized decomposition (PGD) constructs a separated representation of the solution directly from the governing equations, bypassing the need for full-order snapshots and enabling multi-physics ROMs by construction.

Certification and trust: Aerospace certification authorities require evidence that simulation tools are reliable. ROMs, by their approximate nature, must be accompanied by rigorous error estimates. Researchers are developing a posteriori error bounds that quantify the difference between the ROM and the full-order solution for a given input. For safety-critical applications like engine disc burst or combustor casing rupture, these error estimates are essential before ROMs can be accepted in the certification workflow.

Machine Learning and Data-Driven ROM: The Next Frontier

The intersection of reduced-order modeling and machine learning is accelerating progress. Supervised learning methods can train ROMs that map input parameters (geometric dimensions, material properties, boundary conditions) directly to outputs of interest (thrust, specific impulse, outlet temperature) without explicitly constructing the full-order state. Neural network architectures such as convolutional autoencoders and physics-informed neural networks (PINNs) incorporate conservation laws into the training loss, ensuring the ROM respects physical constraints even in extrapolation regions.

For propulsion simulations, one of the most exciting developments is the use of operator learning frameworks like DeepONet and Fourier Neural Operators (FNO). These learn the mapping from parameters to solution fields directly from data. Once trained, an FNO can predict the entire flow field for a new combustor geometry or operating point in under a second, with accuracy that rivals a coarse-mesh CFD solution. Researchers at Ansys and Siemens Digital Industries Software are actively integrating operator learning ROMs into commercial simulation platforms, aiming to give propulsion engineers the speed of correlations with the accuracy of simulations.

Future Outlook: Real-Time Digital Twins for Propulsion

Looking ahead, reduced-order modeling will be a cornerstone of digital twin technology for propulsion systems. A digital twin is a living virtual representation of a physical engine that updates using sensor data from the real asset. ROMs are ideal digital twin engines because they can run faster than real time and be re-calibrated on the fly using streamed measurements. For example, a gas turbine in service on an aircraft could have a ROM-based digital twin that monitors component temperatures, stresses, and efficiency, predicting remaining useful life and alerting maintenance crews before failures occur.

The European Union’s Clean Sky project and NASA’s Transformational Tools and Technologies (TTT) project have both funded initiatives to develop ROM-enabled digital twins for next-generation engines. In the long term, the goal is to automate the entire simulation-to-decision pipeline: a sensor anomaly triggers a ROM-based assessment within milliseconds, which then recommends a control action or a maintenance schedule. Achieving this vision requires continued advances in ROM robustness, uncertainty quantification, and hardware implementation (e.g., FPGA-based ROMs for on-board deployment).

The role of reduced-order modeling in accelerating propulsion system simulations is already transformative. As computational resources become ever more strained by the demand for higher fidelity, ROM offers a pragmatic, and often necessary, path to faster design cycles, deeper physical insight, and real-time digital operations. For engineers and researchers in propulsion, mastering ROM techniques is quickly becoming not just an advantage, but a core competency.