Discovery of Neural Operator Families via Interpretable Meta-Learning
Name
chen-zituo-smme-meche-2026-thesis.pdf
Size
4.53 MB
Format
Adobe PDF
Checksum (MD5)
1a36b45e4c23fc8a3f568e4b5b228146
Author(s)
Chen, Zituo
Advisor(s)
Deng, Sili
Date Issued
February 2026
Publisher
Massachusetts Institute of Technology
Abstract
Learning solution operators of complex dynamical systems has been a central task of science and engineering, yet bottom-up construction is brittle when the governing physics are not fully known yet. We introduce GUIDE (Gauge-Understand-Identify-Decode-Extrapolate), a data-driven, interpretable meta-learning framework that discovers a family of neural-approximated solution operators to meet the need of generalized surrogates for applications such as fast design iterations and online control. Unlike neural operators in need of explicit or implicit autoregressive conditioning to accommodate multiple dynamics and hardly extrapolate, models trained with GUIDE method can generalize well through an emergent and interpretable latent manifold from limited training trajectories and give a single predictor functioning as a family of neural operators. Meanwhile, GUIDE is inherently architecture-agnostic serving as a general learning principle. A variety of neural operators can be interchangeably integrated to tailor the model for extensive applications such as long-horizon rollouts stability and geometry-aware modeling, paving the way for a general solution for neural surrogate modeling.
MIT Department
Massachusetts Institute of Technology. Department of Mechanical Engineering
Terms of Use
In Copyright - Educational Use Permitted
Copyright retained by author(s)
Persistent DSpace Link
