Enhancing Tradespace Interpretability: A Machine Learning Framework with Model Explanations and Prediction Reliability Metrics
Name
locricchio-aloc5-sm-aeroastro-2026-thesis.pdf
Size
34.83 MB
Format
Adobe PDF
Checksum (MD5)
c6859667ebcb99815fc483643b676ac6
Author(s)
LoCricchio, Angelina M.
Advisor(s)
Hastings, Daniel
Palladino, Anthony
Date Issued
May 2026
Publisher
Massachusetts Institute of Technology
Abstract
Modern systems engineering faces a significant challenge: as system complexity increases, the architectural search space expands exponentially, often stretching the computational limits of traditional tradespace exploration. Although tradespace analysis is a useful technique for quantifying stakeholder utility, the time and resources required to exhaustively evaluate massive design spaces remain a bottleneck. This thesis investigates the extent to which Machine Learning (ML) can enhance the tradespace analysis process by providing rapid, statistically-grounded surrogate models.
Using datasets centered on Earth observation and On-Orbit Servicing (OOS) missions, this research develops a Machine Learning framework to predict system utility and cost. To address the "black box" nature of ML, SHAP (SHapley Additive exPlanations) is utilized to identify design patterns and provide architects with transparent, feature-level insights. Furthermore, this thesis proposes a novel reliability metric designed to warn end-users of invalid or unreliable predictions without the computational overhead of traditional confidence intervals.
The results demonstrate that while ML models can significantly accelerate tradespace exploration by requiring less data generation, their effectiveness is highly dependent on proper data preprocessing, what models are used, and the amount of data given. Further, the SHAP explanation tools validate model plausibility, show feature and target dependencies, and give point specific feature impacts. The proposed reliability metric proved effective for tree-based architectures by utilizing tree disagreement to identify point-specific uncertainty, though it remains limited by global error metrics in boosting-based models like XGBoost. Ultimately, this work provides a methodology for integrating ML into the systems architect’s toolkit, balancing computational efficiency with model interpretability and trust.
MIT Department
Massachusetts Institute of Technology. Department of Aeronautics and Astronautics
Terms of Use
This material is declared a work of the U.S. Government and is not subject to copyright protection in the United States.
Copyright US Government
Persistent DSpace Link
