Evolutionary algorithms need explainability for real-world physics optimization

A new preprint from researchers including Helena Stegherr, Michael Heider, and colleagues examines why evolutionary computation often fails to gain traction in real-world physics-based modeling. The paper, submitted to arXiv on May 27, 2026, introduces five distinct physics-informed optimization problems described by domain experts. These range across different physical domains, each with unique constraints and objectives.

The authors found that all domain experts uniformly require two things from evolutionary algorithms: fast convergence to good solutions and some form of explainability to build trust in the search process. However, other requirements vary significantly by problem. The paper notes that while many existing techniques can improve these aspects—such as surrogate models, visualization, or rule extraction—they have rarely been deployed in complex real-world physics settings. This reveals a clear gap between algorithmic research and practical deployment, which the authors argue must be closed to unlock the full potential of evolutionary computation in critical applications like engineering design and scientific simulation.