JIP-2: GPU-Accelerated AI Reconstructs Collapsed Monuments with Jenga Physics

Published on arXiv, a new paper titled 'GPU-Accelerated Inverse Structural Anastylosis from Block Collapse Dynamics' presents JIP-2 (Jenga Inverse Predictor), a deep learning framework that automates the reassembly of collapsed stone structures. Traditional anastylosis relies on archaeologists manually matching fallen blocks — a labor-intensive and subjective process. The authors frame the problem as an inverse prediction task, inspired by the combinatorial complexity of Jenga.

JIP-2 integrates a GPU-accelerated rigid-body physics engine using OBB/SAT collision detection and a Projected Gauss-Seidel contact solver (accelerated with Numba JIT and CuPy CUDA). It applies Ziglar's analytical force thresholds (F_app = 3*mu_s*m*g for Y-axis torque-free, and 4*mu_s*m*g for X-axis torque risk) across three friction levels (mu_s in {0.25, 0.40, 0.60}) over 450 simulated episodes. A dual-stream ResNet-18 is trained with one-hot encoded friction vectors to jointly predict block removal count, per-position probabilities, center-of-mass imbalance, and torque risk. The output includes a smooth 3D video of the reverse reconstruction.

The framework shows promise for real-world conservation, particularly at the UNESCO Maya site of Uxmal, Yucatan, where precise reassembly of fallen structures is critical. The paper provides full technical details on pipeline architecture and loss formulation, and the code is linked via GitHub.