EngiBench: A Framework for Data-Driven Engineering Design Research

Abstract: Engineering design optimization seeks to automatically determine the shapes, topologies, or parameters of components that maximize performance under given conditions. This process often depends on physics-based simulations, which are difficult to install, computationally expensive, and require domain-specific expertise. To mitigate these challenges, we introduce EngiBench, the first open-source library and datasets spanning diverse domains for data-driven engineering design. EngiBench provides a unified API and a curated set of benchmarks -- covering aeronautics, heat conduction, photonics, and more -- that enable fair, reproducible comparisons of optimization and machine learning algorithms, such as generative or surrogate models. We also release EngiOpt, a companion library offering a collection of such algorithms compatible with the EngiBench interface. Both libraries are modular, letting users plug in novel algorithms or problems, automate end-to-end experiment workflows, and leverage built-in utilities for visualization, dataset generation, feasibility checks, and performance analysis. We demonstrate their versatility through experiments comparing state-of-the-art techniques across multiple engineering design problems, an undertaking that was previously prohibitively time-consuming to perform. Finally, we show that these problems pose significant challenges for standard machine learning methods due to highly sensitive and constrained design manifolds.

• The paper introduces EngiBench as an open-source framework that standardizes engineering design problems with a unified API across multiple simulation domains.
• It employs a modular Python API supporting inverse design, surrogate-based optimization, and physics-informed neural networks to streamline complex design tasks.
• Experimental results highlight the framework’s effectiveness in reducing simulation failures and enabling rigorous cross-domain algorithm evaluations.

Summary of EngiBench: A Framework for Data-Driven Engineering Design Research

Introduction

The "EngiBench: A Framework for Data-Driven Engineering Design Research" paper presents EngiBench, an open-source library aimed at facilitating data-driven engineering design through standardized benchmarks and interfaces. Historically, engineering optimization tasks have required significant computational resources due to physics-based simulations, which are challenging to implement and execute. EngiBench addresses these issues by offering a comprehensive library of datasets and problems spanning diverse domains such as aeronautics, heat conduction, and photonics, thereby enabling reproducibility and fair comparison of optimization algorithms.

Figure 1: Overview of EngiBench and EngiOpt components.

EngiBench Architecture and Features

EngiBench standardizes the representation of various engineering design problems, integrating datasets, simulation engines, and visualization tools into a single framework. The library utilizes a modular Python API, allowing for the easy swapping of algorithms and problems. Its key features include:

• Unified API: The API supports inverse design, surrogate-based optimization, and physics-informed neural networks. A sample code usage showcases how interactions with the library are streamlined and intuitive.

from engibench.problems.airfoil.v0 import Airfoil problem = Airfoil() problem.reset(seed=42) generated_design, _ = problem.random_design() opt_design, history = problem.optimize(generated_design, {"mach": 0.3, "reynolds": 1e6})

• Cross-domain Benchmarking: Through standardized metrics and constraint-checking utilities, EngiBench supports rigorous cross-domain algorithm evaluations.

  Figure 2: Visualization of some of the implemented problems.

Implemented Problems and Capabilities

EngiBench accommodates a diversity of real-world engineering design challenges, from aerodynamics in Airfoil tasks to electromagnetic design in Photonics. Each problem encompasses:

• Diverse Design Representations: Solutions range from vector-based to image-based configurations, catering to varying computational resources and expertise levels.
• Multiphysics Modeling: Includes implementations that consider both single-physics (fluid dynamics, thermal compliance) as well as complex multi-physics interactions (thermoelastic structures).

For example, in the Airfoil problem, the objective involves optimizing aerodynamic performance by minimizing drag while matching a lift coefficient constraint. On the other hand, the Photonics2D problem focuses on optimizing the routing of electromagnetic waves through different materials.

Experimental Results

Through proof-of-concept experiments, EngiBench demonstrates striking results in terms of algorithmic comparisons across tasks. Generative models like GANs and diffusion models show varying degrees of success depending on task complexity and physical constraints:

• Airfoil Task: Domain-specific models such as B ezierGAN significantly reduce meshing failures, highlighting the importance of parameterization in complex geometrical designs.

Figure 3: Example outputs of our generative models on the Airfoil problem.

• Surrogate Models: While successful on simpler tasks, surrogate models struggle with fidelity in highly nonlinear environments like PowerElectronics, underscoring the necessity for domain-informed enhancements.

Challenges and Future Directions

While offering valuable capabilities, EngiBench does not yet address unstructured mesh or dynamic simulation scenarios. Additionally, biases inherent in simulation assumptions remain a critical area for further research. Future iterations might include extensions to tackle these limitations, broaden practical applications, and enhance cross-task generalization capabilities.

Conclusion

EngiBench provides a versatile and reproducible platform for advancing engineering design research through its standardized datasets and open-source libraries. By streamlining complex simulation integration and offering diverse benchmarking environments, it opens new avenues for research in engineering optimization and machine learning.

The framework and its components reference various works in the field, notably in aerodynamics, topology optimization, and machine learning methodologies. Further information can be found at the project's documentation website and associated code repositories. This research has been supported by ETH Zurich and the University of Maryland.