Modern product development increasingly requires simultaneous optimization across multiple physics domains, from fluid dynamics and thermal management to structural mechanics and electromagnetics. The complexity of these multi-physics design spaces has grown beyond what traditional CAD-driven workflows and sequential simulation pipelines can efficiently address. Engineering teams face unprecedented challenges as products now span multiple physics and domains, while the number of viable design options far exceeds what can be manually explored within typical development timelines.

Generic large language models (LLMs), while powerful in natural language tasks, lack the spatial and physical reasoning required to generate meaningful engineering geometry or reliably predict performance outcomes across these domains. Traditional parametric optimization approaches, constrained by predefined templates and limited design space exploration, result in sub-optimal concepts that fail to leverage the full potential of modern computational capabilities. The fragmented nature of conventional design processes, where isolated teams work sequentially, creates lengthy back-and-forth cycles that delay time-to-market and prevent the holistic optimization required for next-generation product performance.

This presentation describes a technical approach with an electronics cooling use case to closing this gap through a physics and geometry-aware AI architecture capable of generating CAD-ready 3D geometry directly from high-level design intent. The method integrates spatial reasoning with learned representations of physical behavior, enabling AI-assisted exploration of design spaces that are orders of magnitude larger than those accessible through conventional manual iteration. Unlike black-box optimization tools that remove engineer control, this approach implements a builder-in-the-loop methodology that maintains human oversight while dramatically accelerating the design exploration process.

The architecture combines geometric deep learning models for baseline exploration with localized models for refinement, leveraging non-parametric approaches that can navigate complex, non-linear interactions between design parameters. This dual-model strategy enables efficient exploration of the entire design space while providing targeted optimization within promising regions. For electronics cooling applications, this translates to automated generation and evaluation of thermal management solutions that balance heat dissipation, pressure drop, manufacturing constraints, and geometric packaging requirements simultaneously.

Matt Ellis is the Head of Engineering Enablement at Neural Concept, bringing a career spanning technical leadership, product development, and sales in the engineering software and simulation space. He is passionate about pushing the boundaries of product design and helping engineering teams leverage AI to design, simulate, and optimize faster than ever. Matt is a trusted voice in the CAE and engineering intelligence community, regularly engaging at major industry events and working hands-on with customers to turn AI into measurable industrial impact.