STEP: extraction of underlying physics with robust machine learning.

Area of Science:

• Physics
• Machine Learning
• Computational Science

Background:

• Inverse problems are crucial in modern physics, requiring extraction of physical quantities from experimental data.
• Current end-to-end machine learning methods often need complex estimators to learn underlying physical models.
• This complexity can hinder accuracy and generalizability in data analysis.

Purpose of the Study:

• To present a new machine learning paradigm, Surrogate Training Embedded in Physics (STEP), for tackling inverse problems.
• To demonstrate that by making physical models auto-differentiable, complex physics can be leveraged without relearning.
• To showcase STEP's ability to create accurate and robust neural surrogates for unknown physical quantities.

Main Methods:

• Developed the Surrogate Training Embedded in Physics (STEP) approach.
• Utilized auto-differentiable physics models to construct neural surrogates.
• Applied STEP to dynamic kernel deconvolution for analyzing resonant inelastic X-ray scattering (RIXS) spectra.

Main Results:

• STEP generalizes well and is robust against overfitting and significant data noise.
• Demonstrated successful application of STEP in analyzing RIXS spectra.
• Showcased that simple estimator architectures are sufficient for extracting relevant physical information using STEP.

Conclusions:

• STEP offers an efficient alternative paradigm for solving inverse problems in physics.
• The auto-differentiable physics model integration simplifies neural surrogate construction.
• STEP facilitates accurate and robust extraction of physical information, even with noisy data and simple architectures.