Session: 02-05-01: Artificial Intelligence and Data-Driven Design 2

Paper Number: 198664

Knowledge Graphs for Agentic Systems Design: Demonstration in the Lightweighting of Automotive Components 

Large, complex engineering systems are the consequence of a multitude of decisions made by hundreds of individuals ranging from component geometry and manufacturing methods to material selection and processing conditions. Each of these decisions are tightly coupled for a given component and changing any one of them (ex. material choice) could result in suboptimal component performance or incompatibility with existing manufacturing processes. Furthermore, systems-level decision making is often inadequately informed because specific knowledge of subsystems is typically dispersed and institutional. These challenges significantly limit the ability to optimize such systems. For example, efforts to reduce the weight of a motor vehicle are often challenged by a large number of components. With so many parts contributing to the overall mass, it is difficult to determine which components are worth targeting and which are unlikely to result in meaningful savings. 

In the present work, this challenge is addressed by representing system information in a knowledge graph. The developed knowledge graph connects component designs to feasible manufacturing routes and associated costs. The graph construction process begins with component inputs including the design envelope, operating conditions, and the space of possible material properties which are used to populate a series of parametrized topology optimizations. Topologies are decomposed into geometric features which in turn are connected to specific manufacturing routes, including material, processing condition, and manufacturing steps. Material information contained in the knowledge graph is constructed from a database of over 100 alloy compositions including steels as well as aluminum, titanium, magnesium, and nickel alloys among others. Each entry also contains associated density values and condition-dependent strength values. Over 100 manufacturing methods are also represented in the graph including a variety of forming, machining, welding, casting, forging, and additive manufacturing methods. Each method includes a description and is associated with feedstock inputs (ex. cast ingot, billet, plate, bar, etc.), geometric feature inputs and outputs, and applicable alloys. Additionally, over 150 processing methods are represented in the graph including surface treatment, heat treatment, and as-fabricated conditions for forming, welding, additive manufacturing, etc. 

A case study was developed to demonstrate the effectiveness of the knowledge graph for automotive lightweighting. A reduced set of components were selected from the suspension system (ex. wheel, control arm, etc.) and optimized considering both static and dynamic loads via Altair HyperMesh software. Each component was optimized for a range of stress constraints and material Poisson’s ratios to capture the range of optimal geometries associated with each material. An interpolation scheme was then used to determine the optimal geometry for a given set of material properties, which in turn was used to produce necessary geometric features and resulting component mass.

After the knowledge graph was assembled, an optimization scheme was developed to determine the set of component geometries and associated manufacturing routes (ex. method, material, process) that are Pareto-optimal with respect to weight, cost, and performance. The knowledge graph and associated optimization scheme were integrated with a locally hosted large language model (OpenAI’s gpt-oss-120b). The resulting tool successfully provided accurate responses to a range of natural language queries including lightweighting designs constrained by cost or available manufacturing methods and component prioritization in systems-level optimization exercises.

The present work demonstrates how knowledge graphs with embedded manufacturing information can be used to efficiently explore alternative component designs and manufacturing routes, prioritize components for system-level lightweighting initiatives, and to encode system information in large language models for AI-assisted decision making.

Presenting Author: Tanner Kirk SandboxAQ

Presenting Author Biography: Tanner Kirk is a Staff Research Scientist and the Alloy Lead at SandboxAQ. Tanner leads the development of quantitative models that integrate physics-based techniques with AI and machine learning to predict alloy microstructure and properties. His work at SandboxAQ also focuses on building software that leverages these alloy models and encoded knowledge to guide decision making for metal manufacturing. He holds a Ph.D. in Mechanical Engineering from Texas A&M University, with a specialty in Integrated Computational Materials Engineering.

Authors:

Tanner Kirk SandboxAQ
Maciej Polak SandboxAQ
Bill Feng SandboxAQ
Eric Rothchild SandboxAQ
Maritza Hernandez-Gaete SandboxAQ