SeqGPT: Transformer slashes composite design time from hours to seconds

Composite material design often involves optimizing stacking sequences (the order and orientation of layers) to meet mechanical targets like buckling parameters, all while adhering to discrete manufacturing constraints and global continuity between adjacent panels (blending). Traditional iterative methods are computationally expensive, especially for multi-panel configurations. Now, researchers from Toulouse INP and IRIT present SeqGPT, a conditional Transformer agent trained to predict valid stacking sequences directly. The model uses a neurosymbolic decoding strategy: a Transformer predicts a conditional distribution over possible sequences, and a constrained beam search prunes any branch that violates blending rules, ensuring both feasibility and global continuity by construction.

Tested on the challenging 18-panel horseshoe benchmark, SeqGPT generates solutions near-instantaneously — a stark contrast to hours or days required by evolutionary optimizers. Despite the speed, buckling performance remains comparable to state-of-the-art evolutionary methods. This work (arXiv:2607.11910, presented at APIA 2026) demonstrates how constrained generative models can replace expensive solvers in engineering inverse problems, potentially accelerating composite design workflows in aerospace, automotive, and other high-performance industries where lightweight, strong structures are critical.