What if we could backpropagate through the law? 
This is the research question behind #normax, my entry to the 2026 Tesseract Hackathon by Pasteur Labs.
Engineering lightweight structures remains largely a sequential, disconnected process: from form-finding to establish a structure’s shape, to structural analysis to evaluate its response under multiple load cases, all the way to code verification for safety approval. In this sequence, engineering decisions move forward, but performance feedback seldom travels back.
To address this challenge, #normax turns these three stages of structural engineering into one differentiable program using JAX. Derivatives flow backward from code utilization, through structural response, to the input parameters, enabling fast gradient-based optimization of shape and cross-sections to minimize structural weight.
The gridshell below shows what end-to-end backpropagation can offer when designing a roof under three load cases. On the left is business as usual: pure cross-section sizing under Eurocode 3 rules. Optimizing the gridshell profile, in the middle, reduces mass by 34%. On the right, optimizing form and sections through the full pipeline reaches 41.5% less mass, the lightest of the three designs while satisfying the implemented code checks.
This prototype was built at The Princeton Laboratory for Artificial Intelligence using Tesseract.
Star and peruse the code on GitHub: GitHub - arpastrana/normax: Backpropagating through structural engineering standards. Tesseract Hackathon 2026 submission — Track 01: Inverse Design & Shape Optimization · GitHub