Generating High-Quality Data for Turbomachinery Machine Learning

Turbomachinery optimization has long faced a fundamental tradeoff: the design space is highly complex, and evaluating each design with high fidelity is expensive. In compressors, pumps, and fans, the challenge grows when the goal is not a single operating point, but a multipoint, multi-objective, and often multidisciplinary optimization.

Traditional direct-design methods define blade geometry section by section using parameters such as blade angles, stacking lines, and shape-control curves. This creates many input variables, reduces their physical relevance, and makes design exploration inefficient. Inverse design takes a different approach. Rather than modifying geometry and observing the resulting flow, it starts by prescribing key flow quantities and derives the blade shape from them. When combined with high-fidelity computational fluid dynamics (CFD), this physics-based parameterization can produce consistent machine-learning training data that can be used to train high-fidelity predictors. These predictors enable complex optimization and accelerated design-space exploration with relatively small datasets.

Why Inverse Design Improves Design Exploration

One of the defining advantages of inverse design is its physics-driven approach. Rather than starting with blade geometry and iterating toward a desired performance target, engineers specify the desired flow characteristics, such as swirl distribution and blade loading, from the outset. This provides direct control over key aerodynamic and hydrodynamic outcomes, including pressure rise, power consumption, and torque.

Since the performance objectives are embedded into the design process, sophisticated three-dimensional blade shapes can be generated with only a small number of design variables, simplifying the development of high-performance fans, pumps, and compressors.

Inverse Design V/s Direct DOE (Zhang et al., 2021)

The value of this approach becomes even more evident during design exploration. In a study published in the Journal of Fluids Engineering, researchers compared conventional direct design and inverse design methods for pump impeller optimization. A design of experiments (DOE) based on direct blade parameterization produced a large scatter of results, with many designs missing the target head or exhibiting poor cavitation performance.

When the same analysis was conducted using inverse design using TurboDesign by Advanced Design Technology, the candidate designs naturally clustered around the desired operating conditions. By embedding the required performance characteristics into the design variables themselves, inverse design narrows the search space and enables engineers to focus on refining efficiency and cavitation behavior, leading to faster and more effective optimization (Zhang et al., 2021).

The Role of Fidelity Turbo: Turning Physics-Led Designs into High-Fidelity Data

Physics-driven inverse design is only one part of the equation. To build reliable AI and machine learning models, engineers also need a steady stream of accurate and consistent performance data. Fidelity Turbo provides this critical link by seamlessly connecting inverse design with high-fidelity CFD simulation.

Designs generated in TurboDesign are automatically meshed in AutoGrid, analyzed in Fidelity Turbo, post-processed, and returned to the same project environment. Every design is evaluated using a standardized CFD workflow—with consistent mesh quality, Y-plus targets, solver settings, boundary conditions, and reporting metrics—the resulting dataset remains highly comparable across cases. This reduces numerical variability and improves the quality of the data used for surrogate modeling and optimization.

TurboDesign's integrated Fidelity Turbo workflow, linking TurboDesign by Advanced Design Technology with Cadence Fidelity Turbo for efficient turbomachinery design and simulation

The integrated workflow is also designed for multipoint analysis, enabling engineers to evaluate designs across multiple operating conditions with minimal manual effort. Mesh generation, solver setup, simulation, and post-processing are automated for each operating point, while key performance metrics such as efficiency, pressure ratio, and flow characteristics are collected in a unified reporting framework.

By combining physics-based inverse design with automated high-fidelity CFD, Fidelity Turbo accelerates the creation of robust datasets for AI-driven optimization while ensuring that every design is assessed under realistic operating conditions.

Case Study: Optimizing a Transonic Compressor with Inverse Design

Setting up reactive response surface optimization (RRS) for the transonic compressor

A compelling demonstration of this integrated workflow is the optimization of a transonic compressor stage with a pressure ratio exceeding 4:1. Rather than relying on dozens of geometric variables, the entire optimization problem was defined using just five physics-based design parameters: four controlling the blade loading distribution and one governing wrap-angle stacking.

The objective was to simultaneously maximize efficiency at the design point while improving pressure ratio at both low-flow and high-flow operating conditions, thereby expanding the compressor's stable operating range.

Baseline vs. Optimized Design Results

Using the TurboDesign integrated Fidelity Turbo workflow, 75 design variants were automatically generated, meshed, simulated, and evaluated across three operating points. From the resulting dataset, a surrogate model was trained using only 64 valid CFD cases, demonstrating how inverse design dramatically reduces the dimensionality of the optimization problem while maintaining direct control over the underlying flow physics.

The results highlight the power of combining inverse design with optimization. The optimized compressor delivered improvements in both efficiency and pressure ratio across the entire operating range, including off-design conditions that are traditionally difficult to improve simultaneously. Detailed CFD analysis revealed a noticeably more uniform exit-flow field, indicating reduced aerodynamic losses and improved flow quality through the impeller.

RRS (surrogate model) prediction accuracy vs. CFD data for Optimal Design

Equally significant was the accuracy of the surrogate model itself: predictions of stage pressure ratio matched high-fidelity CFD almost exactly, while efficiency predictions differed by only a fraction of a percentage point. Achieving this level of predictive accuracy with a relatively small dataset enabled the solution of a complex multipoint, multi-objective optimization problem in approximately 25 hours on a modest four-core workstation.

 The study demonstrates how inverse design not only produces higher-performing compressor geometries but also creates a highly structured design space that allows machine learning models to converge quickly and reliably, reducing the computational cost of advanced turbomachinery optimization.

A Practical Path to Turbomachinery Optimization

The broader significance of this inverse design workflow is that it connects three capabilities that are too often treated separately: a physics-led way to define the blade, a high-fidelity way to evaluate it, and a machine-learning framework that can exploit both. Inverse design reduces the dimensionality of the problem while preserving the flow physics. Fidelity Turbo provides the repeatable, high-fidelity CFD backbone needed to generate trustworthy datasets across multiple operating points.

Machine learning then becomes practical not because the industry suddenly has massive data, but because the workflow is engineered to make every simulated case count. The result is a design process that is faster, more robust, and more capable of addressing the real problems turbomachinery engineers face: multipoint targets, conflicting objectives, structural constraints, and the need to reach high-performing solutions without brute-force exploration. For teams pursuing next-generation compressors, pumps, and fans, that combination is increasingly necessary.

References

Zhang, L., Davila, G., & Zangeneh, M. (2021). Multi-objective optimization of a high specific speed centrifugal volute pump using 3D inverse design coupled with CFD simulations. Journal of Fluids Engineering, 143(2), 021202. https://doi.org/10.1115/1.4048292

To learn more, watch the webinar, End to End to Workflow for Faster Smarter Turbomachinery Design, for a deeper look at how inverse design, high-fidelity CFD, and machine learning can accelerate turbomachinery optimization.