Abstract
The goal in system-level design is to generate a diverse set of high-performing design configurations that allow trade-offs across different objectives and avoid early concretization. We use deep generative models to learn a manifold of the valid design space, followed by Monte Carlo sampling to explore and optimize design over the learned manifold, producing a diverse set of optimal designs. We demonstrate the efficacy of our proposed approach on the design of an SAE race vehicle and propeller.
Introduction
Designing complex physical systems, such as a Formula SAE race car, is a multi-faceted challenge. Engineers must navigate a high-dimensional space of parameters—ranging from engine torque and suspension damping to aerodynamic wing angles—while balancing competing objectives like acceleration, mass, and drag. Traditional automated design processes often struggle with three primary issues: the lack of formal mathematical specifications for "realistic" designs, the prohibitive computational cost of high-fidelity multiphysics simulations, and the tendency of optimizers to converge on a single solution rather than providing a diverse set of options for human trade-off analysis.
The DeLPhy Framework
To address these challenges, the authors propose DeLPhy (Design using Learning focused on Physical systems). This framework shifts the optimization task from the raw parameter space to a learned, lower-dimensional manifold.
Stage 1: Manifold Learning and Surrogate Modeling
DeLPhy utilizes a Variational Autoencoder (VAE) trained on a dataset of exemplar designs. The VAE learns a compressed "manifold" where only valid, physically realistic designs reside. Simultaneously, a Specification Network is trained as a surrogate model. This network predicts performance objectives directly from the latent space ($z$), allowing for near-instantaneous evaluation during the optimization phase, bypassing slow "black-box" simulations.
Stage 2: Exploration via Partial Simulated Annealing
Optimization is performed within the latent space using a novel "partial" simulated annealing (PSA) approach. Unlike standard simulated annealing, PSA allows for independent "cooling schedules" for different design objectives. This flexibility enables designers to prioritize specific targets (e.g., safety or minimum speed) while allowing other objectives to explore a broader range of the design space.
Uncertainty Guardrails
A critical component of DeLPhy is its ability to avoid "hallucinations"—designs that the surrogate model predicts will perform well but are physically impossible. By employing Monte Carlo (MC) Dropout in the specification network, the system quantifies uncertainty. If the model is highly uncertain about a design's performance, it indicates the search has moved "off-manifold" into unreliable territory, and the design is discarded.
Experimental Evaluation
The framework was validated across three engineering case studies, most notably a Formula SAE racing vehicle and a propeller design.
The SAE Race Vehicle study involved 39 parameters across 11 subsystems, including the engine, tires, brakes, and suspension. The goal was to optimize for 11 competing objectives, such as total mass, center of gravity height, and corner velocity. The Propeller Design study focused on geometric properties like blade count and pitch, evaluated using the OpenProp physics tool.
Key Results and Insights
DeLPhy demonstrated several advantages over traditional sampling and optimization methods:
- Discovery of High-Performance Trade-offs: In propeller design, the framework identified configurations maintaining >75% efficiency even at low velocities (<4.5 m/s), a region typically difficult to optimize.
- Reliability and Convergence: The uncertainty-aware specification network correctly flagged "unreliable" designs where surrogate predictions diverged from true physics. It successfully identified regions where the underlying physics simulator (OpenProp) would have failed to converge.
- Design Diversity: For the SAE car, DeLPhy produced a "fleet" of diverse, high-performing configurations. This allows engineers to choose between different mechanical strategies (e.g., different engine-suspension pairings) that all meet the required performance thresholds.
Conclusion
DeLPhy provides a robust methodology for system-level design by constraining optimization to a learned manifold of valid designs. While the approach is dependent on the availability of initial "exemplar" data, it significantly reduces the computational burden of design exploration and ensures the generation of diverse, physically realizable solutions. Future work will focus on scaling the framework to higher-dimensional descriptions and integrating manufacturing constraints directly into the manifold learning process.
Summary updated