Machine learning for aircraft engine design offers a way to use historical engine data to support early design decisions. At the conceptual stage, engineers must compare candidate architectures before every performance detail is available. Predictive models can help estimate selected outcomes quickly, providing another tool for design-space explorationânot a replacement for engineering judgment or detailed analysis.
A NASA study explored this approach for axial-compressor turbofan engines, developing models to predict cruise thrust specific fuel consumption (TSFC) and classify engine core size. Its results illustrate both the potential of supervised machine learning and the importance of understanding the data, assumptions, and limits behind a prediction.
Machine learning for aircraft engine design: two prediction tasks
The models drew on a database of 183 engines: 144 manufactured commercial engines and 39 engines examined in earlier NASA aeronautics projects. The commercial-engine data covered designs from the mid-1960s to the mid-2010s, bringing together information from multiple generations of engine technology.
The study treated the two design questions differently because the available data differed. Cruise TSFC was handled as a regression problem: the model estimated a continuous performance value. Core size was handled as a binary classification problem, distinguishing acceptable from unacceptable sizes. Public data did not provide actual core-size measurements for the commercial engines, so the study classified core size using a threshold rather than predicting an exact dimension.
Inputs and model choices
Estimating cruise TSFC
A deep neural network (DNN) was used to identify relationships between engine and flight-condition inputs and cruise TSFC. Inputs included overall pressure ratio (OPR) and bypass ratio (BPR) at sea-level static conditions, sea-level static thrust, cruise Mach number, cruise altitude, and engine technology level, represented by certification year.
Some influential design informationâincluding turbine inlet temperature, turbine cooling, and turbomachinery efficienciesâwas not publicly available. The study used engine technology level as an indicator to account for their relationship with technology advances. This is an important modeling choice: a proxy can make a model possible when direct measurements are unavailable, but it does not turn the missing variables into observed data.
The DNN had six hidden layers, each with six neurons. The model used backpropagation to adjust its internal weights, dropout regularization to reduce overfitting risk, and the Adam optimization algorithm. It was developed in Keras with TensorFlow as the computational backend.
Classifying engine core size
A support vector machine (SVM) classified core size using OPR, BPR, and sea-level static thrust, along with engine technology level. The target categories were defined by high-pressure compressor (HPC) last-stage blade height, represented as h:
- Acceptable: h â„ 0.50 in.
- Unacceptable: h < 0.50 in.
This is a classification output, not a direct measurement or continuous estimate of core dimensions. In a design workflow, that distinction matters: the model answers whether a design falls on one side of the selected threshold, not how large its core should be in exact engineering terms.
Training, validation, and evaluation
The dataset was randomly divided into 137 engines for training and 46 for testing. The training portion was used to develop and tune the models, with six-fold cross-validation providing a preliminary check across different subsets of the training data. After tuning, all 137 training examples were used to build the final models, which were then evaluated on the 46 engines held out from model development.
This separation is central to interpreting predictive performance. Cross-validation helps assess how a model behaves across training-data partitions, while a held-out test set provides a further evaluation using examples not used to fit the final model. The reported results apply to this dataset and setup; they should not be assumed to guarantee the same accuracy for a different engine population or design space.
What the reported results show
In cross-validation, the study reported average prediction accuracy of 97.9% for cruise TSFC, with 3.5% uncertainty, and 97.8% for core-size classification, with 4.3% uncertainty. On the held-out test set, the TSFC model achieved average accuracy of 98.3%; accuracy exceeded 95% for 45 of the 46 engines and was 94.8% for the remaining engine.
The core-size classifier correctly classified all 46 test engines: 40 in the acceptable category and six in the unacceptable category. The study reported 100% overall test-set accuracy, including correct identification of all six engines below the threshold. These are results from a specific, relatively small dataset, not a general performance guarantee. The report also notes that expanding the database could improve TSFC prediction and reduce uncertainty, while limited public engine data presents a challenge.
Practical implications for design teams
For engine developers and system integrators, the key opportunity is using predictive analytics to screen candidate concepts earlier. Estimates of cruise TSFC can help compare alternatives in terms of a fuel-efficiency measure relevant to fuel burn, aircraft range, and operating cost. Core-size classification can provide an early check against a chosen design threshold as teams explore engine architectures.
To apply this approach responsibly, technical teams should consider:
- Data coverage: Confirm that the training examples are relevant to the engines and operating conditions being assessed.
- Input assumptions: Make proxies, such as certification year for technology level, explicit and review their limitations.
- Validation design: Keep evaluation data separate from model fitting and report uncertainty as well as average accuracy.
- Decision role: Use predictions to inform comparisons and prioritize analysis, not as a substitute for detailed engineering evaluation.
Conclusion
Machine-learning models demonstrated a promising way to use historical turbofan data for conceptual-design support, addressing both cruise TSFC regression and threshold-based core-size classification. The reported test results were strong, but their practical value depends on data quality, relevant coverage, transparent assumptions, and careful validation. For engineering organizations, the most useful role for these models is as a rapid screening aid that helps narrow design options while more detailed analysis remains essential.
Access Official Technical PDF