AI-Driven Synthesis for High-Tech System Design: Au... | AI Research

AI-Driven Synthesis for High-Tech System Design: Automating Innovation
Modern high-tech engineering faces a "combinatorial explosion," where the sheer number of possible components, materials, and configurations makes it impossible for humans to manually explore every design option. This paper introduces Automation-in-Design (AiD) and Computational Design Synthesis (CDS) as a new paradigm to solve this. By using deep learning and generative AI, the authors propose a framework that shifts engineering from traditional, sequential, and simulation-heavy processes toward autonomous design systems that can generate and optimize complex architectures with minimal human intervention.

The Challenge of Integrated Design

Engineering complex systems—such as electric vehicle drive-trains—requires solving three interdependent problems simultaneously: selecting the right components (topology), determining their physical dimensions (dimensioning), and designing the control strategies. Traditional methods often address these sequentially, which leads to suboptimal results. The authors argue that a holistic approach, where physical and control systems are optimized together, is essential to unlocking innovative architectures and reducing costs.

How the Framework Works

The proposed CDS framework functions by mapping system functions to specific components and using AI to navigate the resulting design space. The authors utilize two primary AI-driven schemes:

• Iterative AI Optimization: A reinforcement learning (RL) agent selects high-level system topologies, while a physics-based solver (nonlinear programming) handles the continuous variables, such as gear ratios or dimensions. This "solver-in-the-loop" approach ensures that every AI-generated design is physically feasible.

• Generative AI Optimization: Once the iterative process generates enough data, a predictive model—such as a deep Q-network—is trained to predict optimal topologies almost instantaneously based on design constraints, bypassing the need for time-consuming simulations.

Solving Spatial Complexity

Beyond choosing components, the paper addresses the "dimensioning" problem—the physical placement and routing of parts. Because traditional CAD models are too complex for fast, automated optimization, the authors introduce "Maximal Disjoint Ball Decomposition" (MDBD). This method simplifies complex 3D shapes into a collection of spheres, creating a differentiable model that allows AI to perform gradient-based optimization. This enables the system to automatically find efficient placements for components without requiring constant manual CAD iterations.

Performance and Impact

The framework was validated through two case studies: an automotive gearbox-topology optimization and a spatial packaging problem. In the gearbox study, the AI-driven approach reduced evaluation time by three orders of magnitude compared to brute-force methods, while maintaining results within 2% of the theoretical optimum. The spatial packaging study confirmed that the MDBD method could successfully find configurations that closely match analytical benchmarks. These results demonstrate that AI can effectively manage the trade-offs between efficiency, mass, and physical constraints, paving the way for fully autonomous system design.