What is it about?
Artificial intelligence can speed up computer simulations used in science and engineering, but how can we tell whether its predictions are reliable? We show that predictions can look accurate and score well on common tests while still giving poor estimates of physical quantities, such as how much material flows across a boundary. In a study of diffusion and chemical reactions, an AI model designed to preserve the mathematical structure of the underlying physics produced more reliable results. Our findings highlight the need to train and evaluate scientific AI using the physical quantities that matter in practice.
Featured Image
Photo by Susan Wilkinson on Unsplash
Why is it important?
As AI is increasingly used to accelerate scientific simulations, our work exposes a gap between predictions that look accurate and those that support reliable scientific calculations. We show that even perfectly sampled simulation data can lose essential physical information when represented in a commonly used way, so improving training alone may not resolve the problem. By preserving the mathematical structure of the underlying physics, we obtain substantially better agreement with conservation laws and more accurate boundary-flow estimates. These findings could guide the design and evaluation of AI models for scientific and engineering applications.
Perspectives
This paper grew out of my long-term research on variationally correct operator learning and a question that has become central to my work: does the quantity we train an AI model to minimize actually measure what we need scientifically? I believe this question extends well beyond neural operators to AI for science more broadly. Choosing a convenient training objective can make computation easier, but its meaning and limitations need to be understood within the scientific domain. Balancing computational efficiency with scientific reliability is often difficult. I hope this paper encourages researchers to examine that balance carefully and ask what their models’ apparent accuracy really guarantees.
Yuan Qiu
Georgia Institute of Technology
Read the Original
This page is a summary of: Small L2 Error Is Not Enough: Function-Space Consistency for Neural Operators, August 2026, ACM (Association for Computing Machinery),
DOI: 10.1145/3806096.3844867.
You can read the full text:
Contributors
The following have contributed to this page
