GLUE: Generative Latent Unification of Expertise-Informed Engineering Models

TL;DR

GLUE creates a modular framework to coordinate frozen, pre-trained subsystem generators through a system-level latent space, enabling feasible, high-performing, and diverse full-system designs. It contrasts data-driven DD-GLUE and data-free DF-GLUE training on a high-dimensional UAV problem, showing that data-free learning achieves orders-of-magnitude compute savings while maintaining strong feasibility and near-optimality, with explicit diversity control via DPP. Ablation experiments reveal the importance of subsystem model smoothness and the management of equality constraints to avoid mode collapse, while inner gradient solvers and constraint tolerancing mitigate such issues. Overall, GLUE offers a scalable, collaborative path to deploying multi-domain generative design in real-world engineering, with broad potential for reuse of third-party subsystems and faster design exploration.

Abstract

Engineering complex systems (aircraft, buildings, vehicles) requires accounting for geometric and performance couplings across subsystems. As generative models proliferate for specialized domains (wings, structures, engines), a key research gap is how to coordinate frozen, pre-trained submodels to generate full-system designs that are feasible, diverse, and high-performing. We introduce Generative Latent Unification of Expertise-Informed Engineering Models (GLUE), which orchestrates pre-trained, frozen subsystem generators while enforcing system-level feasibility, optimality, and diversity. We propose and benchmark (i) data-driven GLUE models trained on pre-generated system-level designs and (ii) a data-free GLUE model trained online on a differentiable geometry layer. On a UAV design problem with five coupling constraints, we find that data-driven approaches yield diverse, high-performing designs but require large datasets to satisfy constraints reliably. The data-free approach is competitive with Bayesian optimization and gradient-based optimization in performance and feasibility while training a full generative model in only 10 min on a RTX 4090 GPU, requiring more than two orders of magnitude fewer geometry evaluations and FLOPs than the data-driven method. Ablations focused on data-free training show that subsystem output continuity affects coordination, and equality constraints can trigger mode collapse unless mitigated. By integrating unmodified, domain-informed submodels into a modular generative workflow, this work provides a viable path for scaling generative design to complex, real-world engineering systems.

Figures (22)

• Figure 1: Given: Frozen, pre-trained specialized models with latents $z_i$. Contribution: Coordination with GLUE models from system-level latent $\zeta$.
• Figure 2: Monolithic and distributed generative modeling. Approach (b) enables specialized models but requires latent coordination.
• Figure 3: I: Optimization algorithm for creation of dataset of optimized designs ($\mathcal{Z}_{opt}$) (e.g., gradient descent, Bayesian optimization, or evolutionary algorithms). II: Data-driven GLUE models (GAN, VAE, DDPM, ...) trained on large datasets obtained using I (optimization algorithms). III: Data-Free GLUE. Here, gradient descent on a loss $\mathcal{L} = \mathcal{L}_{feas} + \mathcal{L}_{perf} + \mathcal{L}_{div}$ is used to train GLUE model directly to map $\zeta$ to $\mathcal{Z}$.
• Figure 4: Exemplar aircraft designs for methods I, II and III.
• Figure 5: Traversing non-feasible, non-optimal design region to achieve diversity.