Surrogate Modeling and Explainable Artificial Intelligence for Complex Systems: A Workflow for Automated Simulation Exploration

Surrogate models can make AI systems both faster and more transparent: The paper demonstrates that lightweight "emulator" models trained on carefully selected data can approximate expensive simulations while running much faster, enabling both practical deployment and the application of explainability techniques that would otherwise be computationally prohibitive.

Uncertainty quantification is essential for trustworthy AI decisions: The proposed workflow explicitly incorporates rigorous uncertainty measurement alongside predictions, allowing decision-makers to understand not just what an AI system predicts, but how confident it should be in those predictions—a critical safeguard for high-stakes applications.

Explanations should be validated across multiple models: The methodology evaluates whether explanations remain consistent when using different surrogate models; inconsistent explanations signal that the models may be unreliable, providing an automated diagnostic tool to catch when AI systems haven't adequately learned the underlying system before being deployed.

This approach applies broadly across domains: Rather than being limited to one field, the workflow is designed to work across diverse applications—from engineering design to social and environmental modeling—suggesting that these responsible AI practices can be systematically integrated into many different simulation-based decision-making contexts.

Explainability and efficiency are complementary, not competing goals: By using surrogate models, organizations can achieve both faster decision-making and greater transparency, addressing two major obstacles to responsible AI adoption simultaneously rather than forcing tradeoffs between them.