Performance and Explainability Requirements of Evolutionary Algorithms in Real-World Physics-Informed Optimization

Problem
This preprint addresses the gap in the application of evolutionary algorithms (EAs) to complex, real-world optimization problems, particularly in physics-informed contexts. While existing literature often focuses on simplified problems, the authors highlight that practitioners are hesitant to adopt EAs due to concerns about performance and the explainability of results. The paper identifies the need for EAs that not only perform well but also provide insights into their search processes to foster trust among users.

Method
The authors introduce five real-world physics-based optimization problems, each characterized by domain experts. For each problem, they delineate specific performance and explainability requirements for the evolutionary algorithms employed. The paper discusses existing techniques that can enhance the performance of EAs, such as adaptive mutation strategies and hybrid approaches that combine EAs with local search methods. However, the authors note that these techniques have not been widely applied in complex scenarios, indicating a significant opportunity for improvement. The paper does not disclose specific architectures, loss functions, or training compute details, focusing instead on the qualitative aspects of algorithm performance and user trust.

Results
The authors report that domain experts universally expect rapid convergence to high-quality solutions, with varying additional requirements based on the specific optimization problem. While quantitative results are not provided in the abstract, the emphasis on expert expectations suggests a qualitative assessment of algorithm performance. The paper implies that existing EAs often fall short of these expectations, particularly in terms of explainability, which is critical for user acceptance in real-world applications.

Limitations
The authors acknowledge that their findings are based on expert opinions rather than empirical performance data from the application of EAs to the described problems. They do not provide a systematic evaluation of the proposed techniques against established benchmarks, which limits the ability to quantify the effectiveness of their recommendations. Additionally, the paper does not address potential scalability issues or the computational overhead associated with implementing more explainable EAs.

Why it matters
This work is significant as it bridges the gap between theoretical advancements in evolutionary computation and practical applications in physics-based optimization. By emphasizing the dual need for performance and explainability, the authors advocate for a paradigm shift in how EAs are developed and evaluated. This could lead to increased adoption of EAs in critical domains where trust and understanding of the optimization process are paramount. The insights provided may inform future research directions aimed at enhancing the usability of EAs in complex, real-world scenarios.

Authors: Helena Stegherr, Michael Heider, Nils Meyer, Tobias Thummerer, Thomas Wendler, Pierre Aublin, Ennio Idrobo-Àvila, Lars Mikelsons et al.
Source: arXiv:2605.28164
URL: https://arxiv.org/abs/2605.28164v1