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

Helena Stegherr, Michael Heider, Nils Meyer, Tobias Thummerer, Thomas Wendler, Pierre Aublin, Ennio Idrobo-Àvila, Lars Mikelsons, Sebastian Zaunseder, Jörg Hähner · 2026

Read Performance and Explainability Requirements of Evolutionary Algorithms in Real-World Physics-Informed Optimization by Helena Stegherr, Michael Heider, Nils Meyer, Tobias Thummerer, Thomas Wendler, Pierre Aublin, Ennio Idrobo-Àvila, Lars Mikelsons, Sebastian Zaunseder, Jörg Hähner (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2605-28164.txt.

Introduction

Stegherr and colleagues ask what real users want from evolution-style optimizers on physics-based tasks. Across five expert cases, every expert wanted quick arrival at good answers plus some account of how answers formed, while other wants shifted by case. The paper lines up known methods with those wants and shows a rift between lab work and field use.

Catalog record

CREATORHelena Stegherr, Michael Heider, Nils Meyer, Tobias Thummerer, Thomas Wendler, Pierre Aublin, Ennio Idrobo-Àvila, Lars Mikelsons, Sebastian Zaunseder, Jörg Hähner
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE72995
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2605.28164

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Performance and Explainability Requirements of Evolutionary Algorithms in Real-World Physics-Informed Optimization Helena Stegherr, Michael Heider, Nils Meyer, Tobias Thummerer, Thomas Wendler, Pierre Aublin, Ennio Idrobo-Àvila, Lars Mikelsons, Sebastian Zaunseder, Jörg Hähner Abstract . Evolutionary computation offers a variety of tools to solve complex real-world optimization problems. However, research often focuses on smaller, simplified problems and optimization algorithms that sometimes miss expectations in real-world scenarios. Additionally, trust in the applied algorithm and the solutions it provides is often essential in such settings, but requires an understanding of the search process itself. This leads to evolutionary computation often not being seriously considered by practitioners in many application contexts, among them physics-based modeling. In this article, techniques from evolutionary computation are detailed that can alleviate these problems. First, five real-world physics-based optimization problems are introduced and described by domain experts. For each of these, the requirements for the evolutionary algorithm regarding performance and explainability to increase trust and usability are presented. We found that all domain experts expect fast convergence to a good solution and want some explanations for how the results were formed, while other requirements strongly depend on the respective problem. Finally, we present existing approaches that can be leveraged to improve those aspects of evolutionary algorithms but have to our knowledge never been employed in complex real-world scenarios. This implies a gap between both domains that needs to be closed to exploit the full potential of evolutionary computation. 1. Introduction Applying EC ( EC ) techniques in real-world scenarios can come with a variety of decisions that should be made in close collaboration between domain experts of both fields. This includes, among others, the description and formulation of the optimization problem, but also additional requirements regarding the efficiency and explainability of the optimizer ( 63 ) . However, in many cases the application of EC methods to real-world problems is investigated by experts of only one of the two fields. This can lead to either oversimplified problem formulations or the rejection of evolutionary approaches due to lack of knowledge in terms of how to configure those for the required performance and explain their process and results. This paper therefore aims at detailing a suitable approach for utilizing EC methods in real-world scenarios, specifically focusing on physics-informed EC and SciML ( SciML ). Physics-informed EC methods are versatile approaches to solve complex real-world problems where closed-form solutions are unavailable and the solution space cannot be explored within a feasible time or at a reasonable cost. They are, for example, used in the reconfigurable intelligent surfaces codebook compilation ( 65 ) , but can also be combined with simulations to simheuristics that are able to provide robust solutions to setups that include stochastic elements ( 19 ) . Additionally, ML ( ML ) and EC are two scientific sub-areas of AI that have already been examined in combination for interdisciplinary topics in various publications (cf. ( 102 ) ). Moreover, the intersection of physics-informed approaches, scientific computing, and ML is referred to as SciML , a domain where physical information and constraints are combined with large amounts of available data to generate more precise models ( 6 ) . Here again, EC methods prove to be valuable in many scenarios. They can be effectively combined with physics-informed neural networks ( 98 ; 100 ) , for example, by performing neuroevolution ( 87 ) . These developments likewise show an increased demand for explainability due to their commonly strong focus on application in complex scenarios. Especially in real-world applications, practitioners are more likely to adopt a method if they can trust both its behavior and performance ( 35 ; 72 ) . This trust can be gained through detailed explanations of the methods and models, as well as their solutions and predictions. In the SciML domain, this need for explainability is voiced in ( 73 ) . One possible solution revolves around using an EC approach to identify symbolic expressions for poorly explainable ML models ( 68 ) , specifically utilizing genetic programming-based symbolic regression ( 12 ; 69 ) . This can increase the fidelity and interpretability of a physics-informed neural network ( 58 ) . EC approaches themselves, however, also require explanations and recent research is concerned with finding appropriate techniques ( 5 ; 105 ; 94 ) . For investigating the application challenges of EC in this context, we focus on providing detailed examples for the definition of five real-world optimization problems and the requirements stated by the domain experts towards the EC approach. This includes the necessary performance aspects of the evolutionary algorithm, as well as explanations for its search process and results. We summarize and discuss these requirements, and based on this information, we examine existing applications of EC to these or very similar problems to determine if and how any of the requirements are fulfilled. Then, we highlight the techniques from the EC domain that are already available to improve the performance and the explainability to meet the expectations of the problem experts. With this, we provide valuable information for domain experts from both fields. EC practitioners can use these examples to assess common problem formulations and complexities, as well as the requirements and expectations towards the performance and explainability of the optimizer. Problem experts, on the other hand, can use the details on existing approaches towards fulfilling the requirements to inform their decision when considering using an evolutionary algorithm. The paper is structured as follows: Section 2 details the five physics-based optimization problems, SCARA , FPP , shape optimization with DeepONets, EIT , and dynamic PET . It also provides all information on the requirements that are specified towards EC methods in the respective scenario. Section 3 then summarizes these requirements, describes the current usage of evolutionary algorithms for the problems and details all additional existing and still missing strategies that are necessary to fulfill them. Finally, Section 4 concludes the paper and presents possible future work. 2. Problem Statements and Descriptions This section details five physics-informed real-world problems from various domains for which EC methods may offer advantages for finding appropriate solutions. For each of those, domain experts directly describe their requirements regarding the EC method and its respective explainability. 2.1. SCARA The goal of this example use case is to create a high-fidelity simulation model for a SCARA ( SCARA ), drawing with a pen on a piece of paper. A SCARA typically consists of two rotational axes with the same orientation and possibly a third (often translational) axis, compare Fig. 1 . Figure 1. Illustration of the SCARA , consisting of two actuated rotational axes (red), and a pen (blue) at the end-effector to draw on a sheet of paper. The friction force created by the contact between the pen and the paper must be overcome by the two motors that drive the two axes. The farther away the pen is from the mounting point of the robot, the more torque the motors have to apply to overcome static friction. As soon as static friction is overcome, the friction reduces abruptly and the generated torque must be reduced again to avoid overshooting. This characteristic effect is often referred to as stick-slip-friction and is

Read the complete text