Decades of research has focused on how to simulate the fracture of objects into multiple fragments. Relative to this forward problem, the inverse problem of fractured object reassembly---recovering the complete underlying object from its fragments---remains largely open, despite its high-impact applications in fields including reconstruction of broken artifacts, bone fracture reduction, docking of proteins, DNA sequencing, fossil reconstruction, document restoration and forensics, geoscience, and assembly planning in robotics. The state of the art in practice is to manually reassemble fragments using experience and domain knowledge - an expensive and labor-intensive ordeal. By combining cutting-edge research on geometric deep learning from the computer graphics and machine learning communities, together with well-validated physical simulation algorithms from computational mechanics, this project will study the first practical automatic algorithm for reassembling an object from its fragments, and to evaluate the effectiveness of this algorithm on 3D scans of real-world pottery and bone fragments. The core idea of this project is to overcome the dearth of fracture data by joint learning of fracture simulation and fractured object reassembly. The research plan has three thrusts. The first thrust introduces a fast but physics-aware fracture and weathering simulation network that can be trained from a modest number of accurate simulation results. The resulting trained simulator