Research map / Probabilistic and scientific modeling / Differential equations and PDEs
ANTiC: Adaptive Neural Temporal In Situ Compressor
Sandeep Suresh Cranganore, Andrei Bodnar, Gianluca Galletti, Fabian Paischer, Johannes Brandstetter
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use PATS temporal selection instead of dense sampling (ablation) | ANTiC-LoRA · 2D Kolmogorov flows (16.8 GiB) | total compression: 47x -> 435x | Improved |
| use PATS temporal selection instead of dense sampling (ablation) | ANTiC-LoRA · 3D BSSN simulations (4.2 TiB) | total compression: 3744x -> 6807x | Improved |
| use PATS (physics-aware) instead of sparse sampling (ablation) | ANTiC-LoRA · 2D Kolmogorov flows (16.8 GiB) | temporal retention: 20% -> 37% | Improved |
| replace ZFP with neural field continual fine-tuning | PATS · 3D BSSN simulations (4.2 TiB) | spatial compression: 27x -> 471x | Improved |
| replace full fine-tuning with LoRA in CFT (ablation) | ANTiC · 3D BSSN simulations (4.2 TiB) | total compression: 860x -> 6807x | Improved |
| use continual fine-tuning instead of training from scratch (ablation) | Neural field · 2D Kolmogorov flows (16.8 GiB) | mean relative l2: 1.7e-3 -> 6.1e-4 | Improved |
| use continual fine-tuning instead of training from scratch (ablation) | Neural field · 3D BSSN simulations (4.2 TiB) | mean relative l2: 6.2e-4 -> 4.6e-4 | Improved |
| use LoRA rank 64 instead of full FT (ablation) | Neural field · 2D Kolmogorov flows (16.8 GiB) | mean relative l2: 6.1e-4 -> 4.3e-4 | Improved |
| use LoRA rank 8 instead of full FT (ablation) | Neural field · 2D Kolmogorov flows (16.8 GiB) | mean relative l2: 6.1e-4 -> 6.9e-2 | Hurt |
| increase LoRA rank from 8 to 64 (ablation) | Neural field · 3D BSSN simulations (4.2 TiB) | mean relative l2: 6.0e-4 -> 4.6e-4 | Improved |
| use CFT instead of cold start for training (ablation) | Neural field · 3D BSSN simulations (4.2 TiB) | wallclock (s), H200: ~183-187 -> ~21-24 | Improved |
| use LoRA rank 8 instead of full FT (ablation) | Neural field · 3D BSSN simulations (4.2 TiB) | time per snapshot (s): ~162 -> ~131 | Improved |
Closest papers by meaning
- Universal Physics Transformers: A Framework For Efficiently Scaling Neural Operators (NeurIPS 2024 · similarity 0.84)
- Space and time continuous physics simulation from partial observations (ICLR 2024 · similarity 0.84)
- PEINR: A Physics-enhanced Implicit Neural Representation for High-Fidelity Flow Field Reconstruction (ICML 2025 · similarity 0.84)
- Compute-optimal data scaling for neural surrogates via multi-fidelity training (NeurIPS 2026 · similarity 0.83)
- Accelerating Parallel Diffusion Model Serving with Residual Compression (NeurIPS 2025 · similarity 0.83)
- Derived Fields Preserve Fine-Scale Detail in Budgeted Neural Simulators (NeurIPS 2026 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
