3D Geometry Generation for Scientific Computing: Agentic 3D/4D Modeling for Scientific Simulation and Design
OpenReview submission portal: 3D4S at CVPR 2027
We invite research on agents that reconstruct, generate, or refine 3D structures and time-varying 3D models for scientific inference, physical simulation, or function-driven design. Work should identify a scientific objective beyond visual appearance and explain how agent actions or feedback contribute to it.
Topics include measurement-grounded reconstruction, geometry and meshing for solvers, inverse design, dynamic physical identification, scientific tool orchestration, and evaluation under uncertainty. See the full topic list. We welcome rigorous negative results, benchmarks, and reproducible scientific software as well as new methods and applications.
The workshop is non-archival. Accepted papers will be made available through OpenReview with author consent under the announced submission terms; they will not appear in IEEE/CVF proceedings.
Papers will follow the CVPR 2027 submission style and applicable format and length restrictions. Regular papers must contain 4-8 content pages, excluding references and appendices. Submissions must be anonymized, including supplementary material.
All deadlines are at 23:59 Anywhere on Earth (AoE).
Papers receive double-blind review with three independent reviews per paper. Reviewers assess relevance, technical correctness, strength of evidence, reproducibility, and clarity. Each author may be listed on at most ten submissions to this workshop, as required by CVPR 2027. Authors and reviewers must declare conflicts of interest through OpenReview. Non-conflicted coordinators resolve disagreements and request additional reviews when necessary. Organizer-authored papers are handled entirely by independent, non-conflicted external reviewers and a designated external decision maker.
The review pool includes 40 confirmed reviewers: 20 in 3D, 15 in scientific problems, and 5 in general machine learning. Our recruitment target is 60 reviewers by January 2027, with at most three papers per reviewer and additional capacity for emergency assignments. We recruit through social-media calls, a Google Form for expertise and availability, and organizer or program-committee nominations. Reviewers recruited through open channels must have at least one first- or second-author paper at a peer-reviewed computer vision, machine learning, or scientific venue. Organizers screen expertise and conflicts, assign complementary reviewers, and spot-check reviews before decisions.
Authors should report the scientific question, geometric representation, agent actions and tools, physical assumptions, feedback signals, computational cost, human intervention, and failure modes. Claims should distinguish appearance agreement, geometric validity, numerical physical validity, and independent validation. Preliminary work is welcome when claims match the available evidence.
All accepted papers will be offered poster presentations. Three papers will be selected for contributed talks based on research quality and balanced topic coverage. The workshop does not include a competition or challenge.
Document data permissions and licensing, protect sensitive anatomy or environments, and label synthetic measurements. State model limitations and avoid presenting simulated results as experimental validation. Share code, configurations, and agent/tool traces where permitted; describe constraints on release when sharing is not possible.
Participation is in person in Seattle. The half-day program runs from 8:00 am to 12:30 pm or 12:30 pm to 5:00 pm, with recording and remote-access arrangements provided by CVPR. All participants must register in advance. See the official workshop requirements.
The CVPR 2026 call for papers is preserved for historical reference. For this edition, contact 3d4sworkshop@gmail.com.