Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Active Learning Enhanced Surrogate Modeling of Jet Engines in JuliaSim

Anas Abdelrehim, Dhairya Gandhi, Sharan Yalburgi, Ashutosh Bharambe, Ranjan Anantharaman, Chris Rackauckas

TL;DR

The paper addresses the challenge of building highly accurate yet computationally cheap surrogates for a turbofan jet engine to accelerate design optimization. It leverages the Digital Echo surrogate modeling method within JuliaSim combined with an active-learning framework that adaptively downscales a densely sampled flight envelope to produce well-conditioned training data. The resulting surrogate attains approximately $0.1\%$ global relative error across outputs, with over $98.4\%$ of test cases falling within $0.1\%$ RE and up to $99.7\%$ for some outputs. This work demonstrates a scalable path to ultra-accurate surrogates for aerospace systems and points to future work in edge-case handling and ensemble modeling.

Abstract

Surrogate models are effective tools for accelerated design of complex systems. The result of a design optimization procedure using surrogate models can be used to initialize an optimization routine using the full order system. High accuracy of the surrogate model can be advantageous for fast convergence. In this work, we present an active learning approach to produce a very high accuracy surrogate model of a turbofan jet engine, that demonstrates 0.1\% relative error for all quantities of interest. We contrast this with a surrogate model produced using a more traditional brute-force data generation approach.

Active Learning Enhanced Surrogate Modeling of Jet Engines in JuliaSim

TL;DR

The paper addresses the challenge of building highly accurate yet computationally cheap surrogates for a turbofan jet engine to accelerate design optimization. It leverages the Digital Echo surrogate modeling method within JuliaSim combined with an active-learning framework that adaptively downscales a densely sampled flight envelope to produce well-conditioned training data. The resulting surrogate attains approximately global relative error across outputs, with over of test cases falling within RE and up to for some outputs. This work demonstrates a scalable path to ultra-accurate surrogates for aerospace systems and points to future work in edge-case handling and ensemble modeling.

Abstract

Surrogate models are effective tools for accelerated design of complex systems. The result of a design optimization procedure using surrogate models can be used to initialize an optimization routine using the full order system. High accuracy of the surrogate model can be advantageous for fast convergence. In this work, we present an active learning approach to produce a very high accuracy surrogate model of a turbofan jet engine, that demonstrates 0.1\% relative error for all quantities of interest. We contrast this with a surrogate model produced using a more traditional brute-force data generation approach.
Paper Structure (7 sections, 2 equations, 4 figures, 4 tables, 2 algorithms)

This paper contains 7 sections, 2 equations, 4 figures, 4 tables, 2 algorithms.

Table of Contents

  1. Nomenclature & Units
  2. Introduction
  3. Methods
  4. Surrogate Modeling Method - the Digital Echo
  5. Active Learning
  6. Results and Error Statistics
  7. Conclusion

Figures (4)

  • Figure 1: Surrogate model generation pipeline in JuliaSim
  • Figure 2: Distribution of the input parameter space (flight envelope)
  • Figure 3: Output distribution pre and post adaptive downsampling approach. In \ref{['fig:pre_post_al_suba']}, the output distribution exhibits skew, whereas the downsampling scheme results in a more uniform output distribution in \ref{['fig:pre_post_al_subb']}
  • Figure 4: Distribution of error across input space, considering only points with >0.1% relative error for the output quantity PerfInst_Fn. A 3D plot of the error across the flight envelope is shown in \ref{['fig:suba']}, which indicates that most of the error is concentrated at the boundaries. After cutting the input space by 20%, i.e., by reducing the mach number input space from (0.0, 0.5) to (0.0, 0.4), the number of high error points reduces by an order of magnitude, as shown in \ref{['fig:subb']}. Finally, an analyses of the remaining points shows that almost all these remaining points have error between 0.1% and 0.2% relative error.