Session: 01-05-01: Foundations of AI and Machine Learning for Aerospace Structural Analysis

Paper Number: 183437

183437 - A Multi-Agent Ai Framework for Multiscale Simulation of Composite Structures 

In the era of advanced composite materials and rapidly evolving computational workflows, this work presents a multi-agent artificial intelligence (AI) framework embedded into CompositesAI, an AI-powered composite expert system that orchestrates simulation of composite materials and structures. The workflow leverages a suite of well-established engineering tools, including Gmsh, SwiftComp, TexGen, VABS (via its PreVABS interface), and GEBT. The system employs intelligent agents to interpret user specifications, generate and execute workflows, enforce quality checks, and post-process results with minimal human intervention.

Each software module has its own agent and dedicated function calls, providing modularity and design freedom for users. The Gmsh agent manages modeling of the structure genes (SGs), which can be 1D, 2D, or 3D, representing microstructures such as 1D SGs for composite laminates, 2D SGs for unidirectional fiber-reinforced composites, or 3D SGs for particulate inclusions and honeycomb structures. The agent configures meshing strategies, boundary conditions, and other parameters, and then generates a mesh file that can be directly used for subsequent analyses.

The platform includes a dedicated textile-modeling agent that uses TexGen to construct various textile composites with multiple weave patterns, based on geometric inputs provided by the user, handles the generation of mesoscale architecture, and prepares models that can be directly passed to the homogenization workflow.

The SwiftComp agent serves as multiscale modeling engine, built on mechanics of structure genome (MSG). Depending on the type of analysis, users can compute effective thermomechanical properties from the individual constituent properties. The SwiftComp input file can be constructed by either Gmsh or TexGen. In a typical microscale analysis, the microstructure is modeled in Gmsh and subsequently homogenized using SwiftComp to obtain effective properties. For textile composites, the yarn-level architecture is modeled in TexGen and homogenized in SwiftComp to compute effective properties. When desired, the workflow can also be extended to perform hierarchical homogenization using both Gmsh and TexGen, where microstructural properties are first obtained using Gmsh and SwiftComp at the fiber and matrix level, then used as constituent data for TexGen-based mesoscale homogenization at the composite level, all coordinated automatically by the agents.

This multi-agent workflow is extended to handle slender composite beams, which are of particular interest in rotorcraft and wind energy applications. The first stage involves constitutive modeling using 2D cross-sectional analysis. Through the VABS agent, the framework computes the cross-sectional mass and stiffness matrices. These computed properties are then passed to the GEBT agent, which performs global 1D beam analysis to obtain the displacements and rotations along the beam axis. The resulting 1D solution can subsequently be fed back into VABS for dehomogenization, enabling recovery of full 3D field quantities such as displacements, strains, and stresses over the original cross-section. This integrated workflow is demonstrated for a thermomechanical analysis of an initially curved slender composite structure.

The multi-agent architecture offers several key advantages over a single, fixed backend API. The overall task is decomposed into specialized, reusable agents, each responsible for a distinct stage such as geometry generation, meshing, homogenization, or structural analysis. The framework is inherently extensible, because new tools can be incorporated by adding or replacing agents without redesigning the entire pipeline. Agents can respond to intermediate results such as mesh quality indicators or solver convergence issues by adapting parameters, refining meshes, or altering the workflow logic, rather than relying on rigid hard-coded assumptions. Taken together, these features enable workflows that are more flexible, robust, and maintainable, while reducing manual scripting effort and supporting systematic, multiscale composite simulation within a single expert system.

Presenting Author: Avinash Rao Purdue University

Presenting Author Biography: Avinash Ramakrishnan Rao is a graduate researcher and aerospace engineer pursuing his M.S. in Aeronautics and Astronautics at Purdue University, specializing in structures and computational engineering. His research focuses on multiscale composite modeling & AI-driven simulation frameworks.

Before joining Purdue, Avinash worked at Mahindra & Mahindra Ltd. as a Senior Engineer in Brakes and Active Safety Systems, where he contributed to vehicle dynamics and product development initiatives. He earned his B.E. in Mechanical Engineering from Birla Institute of Technology, Mesra, and conducted undergraduate research at the Indian Institute of Science (IISc) on shell modeling techniques.

Avinash is passionate about bridging computational mechanics with intelligent design tools, aiming to advance next-generation aerospace and composite structures through automation, efficiency, and data-driven engineering.