These slides were presented at the NAFEMS World Congress 2025, held in Salzburg, Austria from May 19–22, 2025.

Abstract

The energy required to run AI tasks is already accelerating. The World Economic Forum estimated an annual growth rate between 26% and 36%. The energy consumption poses significant sustainability challenges, as the environmental impact of intensive computational processes continues to grow. This also inevitably impacts engineering design optimization methodologies aimed at achieving optimal design performance within constraints. It often necessitates numerous simulations or experiments which can be computationally intensive and time-consuming, especially for complex systems. AI/ML plays a key role in providing computationally efficient surrogate models that allow for more rapid exploration of design spaces. Therefore, even in the context of engineering design optimization, improving the efficiency of AI/ML models is not just a technical necessity but also an urgent environmental imperative. We propose an approach based on an integrated simulation workflow with Design of Experiments (DOE), Response Surface Models (RSM), and Reduced Order Models (ROM). A key component is the reuse of existing data. By repurposing relevant datasets from previous projects, we can significantly reduce the need for new simulations or experiments, thereby conserving computational resources. This not only enhances sustainability but also accelerates the training and validation process of AI/ML models. Even when data is not available, smart incremental and adaptive Design of Experiments (DOE) strategies help to generate a well-distributed set of designs within the input domain space, capturing the underlying behavior of the system accurately with a minimal number of simulations or experiments. Next, RSM models are built to explore different model types and parameters. This exploration helps in finding the most suitable approximation for specific problems and our integrated approach fosters faster model parameters tuning and cross-validation of candidate models. For more complex systems, we implement ROM techniques based on Proper Orthogonal Decomposition (POD). POD-based ROMs offer significant advantages in dimensionality reduction, crucial for handling high-dimensional data, where the curse of dimensionality poses several challenges. By projecting the full- order model onto a lower-dimensional subspace, POD-ROMs drastically reduce computational requirements while maintaining accuracy. Furthermore, POD-based ROMs enhance the interpretability of AI/ML models. By identifying the most significant modes of the system, POD provides insights into the underlying physics, making the AI/ML predictions more transparent and easier to validate. This interpretability is crucial in engineering applications where understanding the model's decision-making process is as important as its accuracy. The integration of DOE, RSM, and ROM into an automated simulation workflow accelerates various tasks, including design exploration, sensitivity analysis, and optimization calculations. We demonstrate the efficacy of this approach in optimizing business jet performance, by exploiting a multidisciplinary approach that encompasses both fluid dynamics and electromagnetics, along with their interactions. This study was made possible only by using AI/ML methods. The proposed methodology successfully addressed the efficiency challenges in AI/ML models through several mechanisms: 1. Data reuse and smart DOE reduce the need for extensive new data generation. 2. Workflow automation enhances RSM model parameter tuning and validation. 3. POD-based ROMs offer dimensionality reduction, crucial for handling high-dimensional data, thereby reducing memory requirements and computational complexity. 4. The enhanced interpretability of POD-ROMs allows for more efficient model refinement and validation, reducing the need for extensive trial-and-error processes in AI/ML model development. In conclusion, this paper presents a comprehensive approach to enhancing AI/ML efficiency in multidisciplinary engineering design optimization. By leveraging DOE, RSM, ROM, and data reuse strategies, we demonstrate significant improvements in computational efficiency, addressing both the technical and environmental challenges posed by the increasing energy demands of AI tasks.