Research map / Video / Video segmentation and tracking

Interactive 4D Volumetric Liquid Forecasting under Moving-Solid Interaction

NeurIPS 2026 · Video › Video segmentation and tracking · approach: Physics-inspired motion prediction

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces interactive 4D volumetric liquid forecasting conditioned on moving-solid trajectories and proposes TIDE/TIDES, interface-driven recurrent forecasters evaluated on a DualSPHysics-generated benchmark.

Problem

Forecasting the future dense 3D Eulerian liquid velocity field over the full domain given the current liquid state and a known future moving-solid trajectory, while remaining stable under autoregressive rollout and distribution shift.

Approach

Generate solid-induced latent updates localized at the evolving solid-liquid interface (TIDE), then add particle-based transport feedback (dual-space P2G) and latent/output correction stages for stabilized rollout (TIDES).

Key results

• TIDES achieves l_roll 0.005746 and traj_end 0.104397, the best among compared models, versus 0.006542 / 0.112111 for U-NO and 3.767 / 5.555 for GNS.
• Response energy drops about 455x from the nearest interface band (1.663763) to the far field (0.003659), motivating interface-localized modeling.
• TIDES inference takes 302.9 ms for an 8-step rollout versus about 6.46 s DualSPHysics simulation time per physical second.

Surprising findings

• Removing dual feedback sharply worsens transport endpoint error (0.104 to 0.154), and removing both corrections causes a larger transport failure, while removing latent correction alone barely changes rollout error.

Limitations

• Synthetic data only; Eulerian target is constructed from particle simulations rather than solver-native field.
• OOD axes limited to motion and object-size shifts.
• Relies on known future solid trajectory and no multi-solid interactions.

Experimental setup

Datasets: DualSPHysics-generated moving-solid liquid benchmark (sphere, cube, cone families)

Benchmarks: moving-solid liquid forecasting benchmark (in-distribution test, motion OOD, size OOD, horizon 16/24)

Metrics: rollout field error near-interface field error far-field error trajectory endpoint error divergence delta

Baselines: 3D ResUNet U-NO Transolver GNS Bulk Only Finite Support

Builds on: U-NO Geo-FNO GNS MeshGraphNets DualSPHysics

Closest papers by meaning

• DeepLag: Discovering Deep Lagrangian Dynamics for Intuitive Fluid Prediction (NeurIPS 2024 · similarity 0.84)
• Bubbleformer: Forecasting Boiling with Transformers (NeurIPS 2025 · similarity 0.84)
• Space and time continuous physics simulation from partial observations (ICLR 2024 · similarity 0.84)
• Gaussian Splashing: Unified Particles for Versatile Motion Synthesis and Rendering (CVPR 2025 · similarity 0.84)
• FluidNexus: 3D Fluid Reconstruction and Prediction from a Single Video (CVPR 2025 · similarity 0.83)
• HelmFluid: Learning Helmholtz Dynamics for Interpretable Fluid Prediction (ICML 2024 · similarity 0.83)