e-SimFT: Alignment of Generative Models with Simulation Feedback for Pareto-Front Design Exploration

TL;DR

e-SimFT introduces simulation-guided fine-tuning (SimFT) to align generative design models with engineering requirements and uses epsilon-sampling to construct high-quality Pareto fronts. It extends this framework to new, unseen requirements via two-step fine-tuning: SFT with simulation data followed by either Direct Preference Optimization (DPO) or PPO with a simulator, enabling principled multi-objective exploration. The approach demonstrates superior Pareto-front hypervolumes compared to baseline multi-objective alignment methods and standard sampling, validated on GearFormer with a physics simulator. This work provides a scalable, simulator-based alternative to human feedback for design exploration and has practical implications for rapid, trade-off-aware engineering optimization.

Abstract

Deep generative models have recently shown success in solving complex engineering design problems where models predict solutions that address the design requirements specified as input. However, there remains a challenge in aligning such models for effective design exploration. For many design problems, finding a solution that meets all the requirements is infeasible. In such a case, engineers prefer to obtain a set of Pareto optimal solutions with respect to those requirements, but uniform sampling of generative models may not yield a useful Pareto front. To address this gap, we introduce a new framework for Pareto-front design exploration with simulation fine-tuned generative models. First, the framework adopts preference alignment methods developed for Large Language Models (LLMs) and showcases the first application in fine-tuning a generative model for engineering design. The important distinction here is that we use a simulator instead of humans to provide accurate and scalable feedback. Next, we propose epsilon-sampling, inspired by the epsilon-constraint method used for Pareto-front generation with classical optimization algorithms, to construct a high-quality Pareto front with the fine-tuned models. Our framework, named e-SimFT, is shown to produce better-quality Pareto fronts than existing multi-objective alignment methods.

Figures (6)

• Figure 1: Randomly sampling a generative model (for engineering design) may not yield a good Pareto front with respect to the design requirements of interest. Also, an engineer could be interested in a new design requirement that the current model is not conditioned on to generate design solutions. We address these issues with SimFT methods -- using simulation feedback to fine-tune a generative model with respect to specific design requirements, including new ones not seen by the model, and proposing a new sampling strategy inspired by the epsilon-constraint method to create a high-quality Pareto front.
• Figure 2: SimFT data generation and training methods.
• Figure 3: Epsilon-sampling with a SimFT model. The target requirement values for $R_1$ is incremented with $\epsilon_i$ and a SimFT model for $R_1$ is sampled to construct a Pareto front.
• Figure 4: Example of gear designs produced for a sample problem with the original GearFormer versus a SimFT model fine-tuned for bounding box volume. The first design has a volume of 0.018$m^3$ while the second design has a much lower volume of 0.008$m^3$.
• Figure 5: Pareto fronts generated by e-SimFT versus other methods for sample design problems.