AI-ML FEM Proposal

This Small Business Innovation Research (SBIR) proposal outlines Andromeda System Incorporatedís (ASI) first phase development of an advanced Artificial Intelligence/Machine Learning (AI/ML) based agent designed to fully automate the generation, prediction, and optimization of finite element models (FEMs) for aerospace structural applications. The Phase I effort will demonstrate the feasibility of finite element mesh generation utilizing our AI/ML-powered toolkit.ÝThe proposed AI/ML innovative tool leverages a multi-modal large language model (MLLM) architecture, such as LLaMA3, paired with a Variational Autoencoder (VAE) to form a two-stage AI system that automates geometry segmentation and mesh seeding. This system will enable automatic generation of meshing-ready surface segmentations and optimized seed points from 2D CAD-derived inputs. The model will be trained on a robust and augmented dataset of annotated aerospace-relevant geometries and evaluated using established mesh quality metrics including aspect ratio, skewness, and Jacobian checks.ÝKey technical objectives include:Defining a dual-module AI framework (segmentation/seeding and mesh correction).Implementing a prototype segmentation and seeding module powered by an 8B-scale MLLM.Developing a training algorithm with expert-annotated and augmented CAD datasets.Establishing an inner-loop training strategy with segmentation-aware loss functions and an outer-loop feedback mechanism driven by mesh quality metrics.Demonstrating feasibility on benchmark aerospace FEM problems using open-source meshing tools.This toolkit will substantially reduce the time, effort, and specialized expertise required to produce high-quality meshes, leading to:Significant reductions in meshing time across a wide range of geometries.Improved mesh consistency independent of user experience.Predictive refinement capabilities without extensive trial-and-error.Fewer simulation reruns, lowering total project costs and time-to-deployment.Phase I deliverables will include a functional segmentation/seeding model, a validated dataset, a mesh quality evaluation framework, and a comprehensive report documenting technical feasibility and readiness for Phase II. Long-term, the system will be designed to integrate into secure, standalone environments under future commercial and government deployment constraints.