AI for Physics and Physics for AI

In the past decade, AI has transformed science and society, with many spectacular success stories, culminating in the 2024 Nobel Prizes in Physics and Chemistry. Beyond the hype and allegations of radical empiricism, the ultimate question is whether Machine Learning (ML) algorithms, regardless of their empirical nature, succeed in capturing levels of complexity that would otherwise remain unattainable by any other methods, including our most powerful theories and computer simulations. Is ML success occasional, or can it become systematic through the development of a comprehensive theory of machine learning?

Despite numerous achievements, several unresolved issues remain:

· How do ML algorithms manage to escape overfitting?

· How does the loss minimization process find useful local minima in high-dimensional rugged landscapes?

· How can we perform robust uncertainty quantification with billions of parameters?

· How can ML be made causal and explainable?

· What are the social implications of ML?

We propose that a satisfactory solution can only result from tight cooperation between theory, computer simulations, and ML.

We invite submissions that explore whether ML's successes are sporadic or systematic, and if a comprehensive theory of machine learning is achievable. Researchers are encouraged to submit papers exploring the theoretical foundations of machine learning, innovative approaches to overcoming overfitting, methods for robust uncertainty quantification, and frameworks for enhancing causality and explainability in AI systems.

Interdisciplinary approaches that bridge theory, simulations, and machine learning are particularly welcomed.

Please note this Research Topic is open to invited contributions only.