Generative AI for Improved Positioning of Human Body Models

Sammanfattning

Finite element human body models (HBMs) are difficult to reposition efficiently for crash and safety studies because practical workflows still involve repeated adjustment, quality checking, and repair. This thesis addresses this workflow-efficiency problem with a hierarchical, constraint-aware artificial intelligence (AI) pipeline for finite element (FE) HBM positioning. The goal is to shorten posture generation while retaining rigid-bone consistency, reconstruction accuracy, and Hex8/Tet4 element-quality checks. Unlike surface-oriented human-body models, the proposed pipeline operates directly on FE nodal coordinates. An anchor multi-layer perceptron (AnchorMLP) predicts rigid bone anchors, Kabsch alignment reconstructs full bone nodes as a non-trainable geometric constraint, and Graph Sample and Aggregate (GraphSAGE) recovers soft tissue with Hex8/Tet4 Jacobian and edge-length penalties. This makes rigid-bone consistency and element-quality control part of generation rather than only post-hoc screening. Collision, range-of-motion, and full contact-screening terms are retained as future evaluation extensions rather than completed quantitative results. The pipeline achieves a Stage 1 bone-node test root mean squared error (RMSE) of 3.26 mm. In Stage 3, Jacobian-aware GraphSAGE reduces soft-tissue active validation mean squared error (MSE) and gives a modest Tet4 quality gain, lowering the best-checkpoint Tet4 low-quality element fraction from 13.5% in the no-Jacobian baseline to 12.7%. End-to-end CPU inference runs in seconds per posture, compared with the hours often required by simulation-driven positioning workflows. Overall, the thesis shows that a three-stage decomposition can support faster positioning, stable skeletal structure, and measurable Tet4 mesh-quality gains. The main limitations are evaluation on one PIPER six-year-old child HBM, residual Hex8 low-quality elements at large joint deflections, and the absence of contact diagnostics and solver-acceptance testing.