Elliptic PDE learning is provably data-efficient.

Area of Science:

• Physics and Machine Learning
• Computational Mathematics

Background:

• Partial differential equations (PDE) learning combines physics and machine learning to identify physical systems from experimental data.
• While current deep learning models excel with limited data, their success is largely empirical.
• Existing PDE learning methods lack theoretical guarantees on data requirements.

Purpose of the Study:

• To provide theoretical guarantees on the number of training pairs needed for PDE learning.
• To develop a provably data-efficient algorithm for recovering solution operators of PDEs.
• To establish an exponential convergence rate for PDE learning algorithms.

Main Methods:

• Exploitation of randomized numerical linear algebra techniques.
• Application of established partial differential equations (PDE) theory.
• Derivation of a novel algorithm for solution operator recovery from input-output data.

Main Results:

• A provably data-efficient algorithm for learning three-dimensional uniformly elliptic PDEs.
• Demonstration of an exponential convergence rate of the error with respect to training dataset size.
• Achieved exceptionally high probability of success in recovering unknown physical systems.

Conclusions:

• Theoretical guarantees can be established for data efficiency in PDE learning.
• Randomized linear algebra and PDE theory enable provably efficient learning algorithms.
• The developed algorithm offers a significant advancement in data-efficient scientific machine learning.