LLM2TEA: An Agentic AI Designer for Discovery with Generative Evolutionary Multitasking

TL;DR

LLM2TEA tackles a key challenge in AI-assisted design by merging LLM-guided multitask evolution with a text-to-3D generator, vision-language evaluation, and physics simulation. It introduces GEM-based agentic design operators that enable cross-domain knowledge transfer and the emergence of novel hybrids beyond single-domain prompts. Empirical results show substantial gains in design diversity (97%–174% relative to a baseline) and physical performance, with over 73% of designs outperforming the baseline mean and several designs being fabricable via 3D printing. The work demonstrates a scalable, practical pathway for AI-driven design optimization and rapid prototyping in engineering contexts.

Abstract

This paper presents LLM2TEA, a Large Language Model (LLM) driven MultiTask Evolutionary Algorithm, representing the first agentic AI designer of its kind operating with generative evolutionary multitasking (GEM). LLM2TEA enables the crossbreeding of solutions from multiple domains, fostering novel solutions that transcend disciplinary boundaries. Of particular interest is the ability to discover designs that are both novel and conforming to real-world physical specifications. LLM2TEA comprises an LLM to generate genotype samples from text prompts describing target objects, a text-to-3D generative model to produce corresponding phenotypes, a classifier to interpret its semantic representations, and a computational simulator to assess its physical properties. Novel LLM-based multitask evolutionary operators are introduced to guide the search towards high-performing, practically viable designs. Experimental results in conceptual design optimization validate the effectiveness of LLM2TEA, showing 97% to 174% improvements in the diversity of novel designs over the current text-to-3D baseline. Moreover, over 73% of the generated designs outperform the top 1% of designs produced by the text-to-3D baseline in terms of physical performance. The designs produced by LLM2TEA are not only aesthetically creative but also functional in real-world contexts. Several of these designs have been successfully 3D printed, demonstrating the ability of our approach to transform AI-generated outputs into tangible, physical designs. These designs underscore the potential of LLM2TEA as a powerful tool for complex design optimization and discovery, capable of producing novel and physically viable designs.

Figures (10)

• Figure 1: LLM2TEA is a novel LLM-driven MultiTask Evolutionary Algorithm for complex design and discovery. LLM2TEA optimizes multiple tasks simultaneously, leveraging cross-domain knowledge learning to drive LLM in evolving prompts. The prompts serve as guides for an auxiliary text-to-3D generative model, expanding exploration into cross-domain regions.
• Figure 2: Instruction sets for the LLM to function as an evolutionary optimizer. LLM2TEA enables the LLM to initialize and evolve prompts guided by various physical and visual criteria in the evolution process.
• Figure 3: Examples of designs generated by text-to-3D generative model are shown. In contrast to the vehicle body shapes generated directly by the model without the LLM2TEA (Fig. \ref{['fig:designs']}(a) and (b)), the designs discovered with LLM2TEA deviate from these conventional shapes, since the model are guided to synthesize designs that not only satisfy physical criteria but also embody the visual characteristics of a car or an airplane, as illustrated by some of the representative example designs in Fig. \ref{['fig:designs']}(c) and (d). The cross-domain search enabled by the multitask setting facilitates the discovery of hybrid vehicle designs, as most evident in Fig. \ref{['fig:designs']}(d).
• Figure 4: Test Scenario I compares between baseline designs generated by the baseline text-to-3D generative model without LLM2TEA to designs discovered with LLM2TEA during the last five generations of a single evolutionary run. The designs discovered with LLM2TEA demonstrated superior expected aerodynamic performance (exhibiting higher normalized lift ($\tilde{C}_{l}$) and lower normalized drag ($\tilde{C}_{d}$) scores for airplanes, as well as lower $\tilde{C}_{l}$ and $\tilde{C}_{d}$ scores for cars) and are visually more novel than the baseline designs. (Note that the illustrations presented here are not drawn to scale.)
• Figure 5: Examples of text prompts (right) used to generate the designs (left) are shown, with colors indicating the magnitude of the fluid velocity, where brighter colors correspond to higher velocity. When LLM2TEA evaluates using only visual criteria, the resulting vehicles tend to have rectangular shapes and high roofs, which correspond to poor drag performance, as illustrated in Fig. \ref{['fig:scientific']} (a). Conversely, relying solely on the physical criterion will degrade the search performance with more highly fragmented vehicle designs discovered, such as those shown in Fig. \ref{['fig:scientific']} (b). As such, by considering both physical and visual criteria simultaneously, LLM2TEA can discover practical designs (Fig. \ref{['fig:scientific']} (c)) featuring body shapes that yield better drag performance compared to baseline (Fig. \ref{['fig:scientific']} (d)).