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

Problem Statement

The application of Evolutionary Computation (EC) in real-world, physics-informed scenarios often faces a disconnect between algorithmic research and practical engineering needs. While EC offers robust tools for complex optimization, practitioners frequently reject these methods due to a lack of trust, insufficient explainability, and the perception that research focuses on simplified benchmarks rather than realistic constraints. This paper addresses the gap between theoretical EC capabilities and the specific performance and explainability requirements demanded by domain experts in physics-informed Artificial Intelligence (AI) and Scientific Machine Learning (SciML). The core problem is defining what is necessary to make EC algorithms trustworthy and usable in high-stakes, physics-based optimization contexts.

Methodology

The authors adopt a qualitative and analytical approach, structured around five distinct real-world physics-based optimization problems. For each problem, domain experts from the respective fields (robotics, composite manufacturing, structural mechanics, medical imaging, and nuclear medicine) were consulted to define:

1. Problem Formulation: A detailed mathematical and physical description of the optimization task.
2. Specific Requirements: Explicit demands regarding the evolutionary algorithm's performance (e.g., convergence speed, global vs. local optimality, constraint handling) and explainability (e.g., trust in results, understanding the search process).
3. Gap Analysis: A comparison of these requirements against current state-of-the-art EC applications in similar or identical domains to identify missing techniques or unexplored strategies.

The five case studies analyzed are:

• Selective Compliance Assembly Robot Arm (SCARA): Hybrid modeling of stick-slip friction using a physics-based model augmented by a feed-forward Artificial Neural Network (ANN).
• Fiber Patch Placement (FPP): Optimization of position and orientation for anisotropic composite patches to maximize structural stiffness under strength constraints.
• Shape Optimization with DeepONets: Minimizing maximum deviatoric stress in a plate with a parametric hole, utilizing Deep Operator Networks (DeepONets) as mesh-less surrogates.
• Electrical Impedance Tomography (EIT): Reconstructing internal electrical conductivity distributions from boundary voltage measurements, an ill-posed inverse problem.
• Dynamic Positron Emission Tomography (dPET): Voxel-wise curve fitting for tracer kinetic modeling to derive physiological parameters from noisy, time-resolved imaging data.

Key Contributions

The paper provides a comprehensive taxonomy of requirements for EC in physics-informed settings and maps existing EC techniques to these needs.

1. Diverse Performance Requirements

The analysis reveals that "one-size-fits-all" performance metrics are insufficient. Requirements vary significantly by domain:

• Optimality: While some problems (SCARA, dPET) accept a "good enough" local optimum due to the oversizing of neural networks or noise, others (FPP, Shape Optimization, EIT) explicitly require robust convergence to the global optimum or highly reproducible results.
• Convergence Speed: All five domains demand fast convergence. This is driven by the high computational cost of evaluating the objective function (e.g., solving ODEs, Finite Element Methods) or the need for real-time processing (e.g., lung monitoring in EIT).
• Constraint Handling: All cases require handling constraints beyond simple boundaries. This includes hard constraints (e.g., patch placement within a domain, stress limits) and soft constraints (e.g., regularization, manufacturing smoothness), necessitating specific constraint-handling strategies.

2. The Critical Role of Explainability

The authors identify explainability as a prerequisite for practitioner adoption. The nature of required explanations varies:

• Search Process: For FPP, Shape Optimization, and EIT, experts require explanations of how the algorithm reached a solution, including internal dynamics and the influence of algorithmic parameters.
• Result Interpretation: For SCARA, FPP, EIT, and PET, explanations of the final solution are vital. This includes understanding the influence of individual variables on the outcome and the stability of the solution across multiple runs.
• Search Space Coverage: Experts in SCARA and PET are interested in knowing how much of the search space was explored to assess the reliability of the found minimum.

3. Identification of Research Gaps

The paper highlights a significant gap: while techniques exist to address these requirements, they are rarely applied or evaluated in complex, real-world physics-informed scenarios.

• Existing Techniques: The authors catalog available EC methods such as Evolution Strategies (CMA-ES) for global optimization, Bayesian statistics for robustness, transfer learning for convergence speed, and various constraint handling methods.
• Explainability Tools: Techniques like Search Trajectory Networks (STNs), behavior analysis, sensitivity analysis, and surrogate modeling for variable contribution are identified as available but underutilized in these specific domains.
• The Gap: The primary missing link is the application and validation of these advanced techniques in complex, real-world settings. Furthermore, there is a need to determine which explanations are actually useful to domain experts versus optimization researchers.

Results and Findings

• Universal Need for Trust: In all five scenarios, the lack of explainability and trust is a primary barrier to EC adoption. Practitioners often prefer less efficient but more interpretable methods (e.g., gradient-based or analytical approaches) if EC cannot justify its results.
• Hybridization is Key: The paper suggests that hybrid approaches (e.g., combining EC with gradient-based methods or using EC to train neural operators) are promising but require careful configuration to meet specific domain constraints.
• Context-Dependent Requirements: The study demonstrates that requirements are not static. For instance, while global optimality is critical for FPP, it is less critical for SCARA where the parameter space is small and multiple acceptable solutions exist.
• Underutilization of Advanced EC: Current research in these domains often focuses on basic performance metrics (e.g., "does it converge?") while ignoring advanced features like search space coverage analysis, robustness quantification, and detailed process explainability.

Significance and Claims

The paper claims to provide a foundational framework for bridging the gap between EC researchers and domain experts in physics-informed optimization. Its significance lies in:

1. Defining the "Real-World" Standard: It moves beyond benchmark problems to articulate the specific, often conflicting, requirements of actual engineering and scientific applications.
2. Bridging the Trust Gap: By explicitly linking explainability techniques (like STNs and sensitivity analysis) to the trust requirements of domain experts, the paper offers a pathway for EC to be seriously considered in high-stakes fields like healthcare and manufacturing.
3. Call for Collaboration: The authors argue that the full potential of EC in SciML cannot be realized without close collaboration. EC practitioners must understand the specific constraints and explainability needs of the problem domain, while domain experts must be informed about the capabilities of modern EC techniques.

The paper concludes modestly, stating that it does not solve these problems directly but rather outlines the necessary requirements and available tools. It posits that future work must involve applying these state-of-the-art techniques to the described problems to validate their efficacy in satisfying expert expectations.