Research map / Architectures and efficiency / Attention and transformers
SCENT: Robust Spatiotemporal Learning for Continuous Scientific Data via Scalable Conditioned Neural Fields
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.
Proposes SCENT, a transformer encoder-processor-decoder conditioned neural field with learnable queries, sparse attention and warp-unrolling forecasting for joint reconstruction, interpolation and forecasting of sparse, noisy scientific spatiotemporal data.
Problem
Spatiotemporal learning from irregular, sparse, noisy, and high-volume scientific data, requiring continuous-in-space-and-time representations that support reconstruction, interpolation, and forecasting at scale.
Approach
A conditioned neural field built on an encoder-processor-decoder transformer with trainable latent queries, sparse self-attention in encoder/decoder, a time-warp processor that conditions on continuous time offsets, and warp-unrolling forecasting that jumps to the time horizon to limit error accumulation.
Key results
- On simulated S1 forecasting, SCENT reaches Rel-MSE 2.51e-2 vs 4.27e-2 for FNO (best baseline), and wins on S2-S5.
- On Navier-Stokes NS-3, SCENT achieves MSE 7.78e-5 vs 1.32e-4 for AROMA; on NS-4 1.03e-1 vs 1.05e-1 and NS-5 1.17e-1 vs 1.24e-1.
- On AirDelhi AD-T and AD-F, SCENT achieves RMSE 53.24 and 45.35, best among baselines; on AD-B AROMA (40.78) beats SCENT (44.20).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add Context Embedding Network (sparse self-attention in encoder) (ablation) | SCENT · S5 ablation | Rel-MSE: +23.3% Rel-MSE when removed | Helped |
| add Calibration Network (sparse self-attention in decoder) (ablation) | SCENT · S5 ablation | Rel-MSE: +11.0% Rel-MSE when removed | Helped |
| add linear projection of input samples (ablation) | SCENT · S5 ablation | Rel-MSE: +4.9% Rel-MSE when removed | Helped |
| provide output time t_o to the encoder (time-targeted spatial encoder) (ablation) | SCENT · S5 ablation | Rel-MSE: +4.29% Rel-MSE when removed | Helped |
| remove all four components (CEN, CN, Proj, TT) (ablation) | SCENT · S5 ablation | Rel-MSE: +67.8% Rel-MSE | Hurt |
| SCENT architecture with warp-unrolling forecasting and time-continuity versus baselines | SCENT · simulated S1-S5 forecasting | Rel-MSE: S1 2.51e-2 vs FNO 4.27e-2 | Helped |
| SCENT vs baselines on AirDelhi AD-B | SCENT · AirDelhi AD-B | RMSE: 44.20 vs AROMA 40.78 | Hurt |
| SCENT vs baselines on AirDelhi AD-T and AD-F | SCENT · AirDelhi AD-T / AD-F | RMSE: AD-T 53.24 vs 55.04 (FNO); AD-F 45.35 vs 47.49 (AROMA) | Helped |
| SCENT vs baselines on NS-4 and NS-5 long-term forecasting | SCENT · NS-4 / NS-5 | Rel-MSE: NS-4 1.03e-1 vs AROMA 1.05e-1; NS-5 1.17e-1 vs 1.24e-1 | No clear effect |
| increase model size (latent dim 128 to 1024) with scaling | SCENT vs FNO · S5 scalability | Rel-MSE: both decrease with size; SCENT follows a linear trend and stays below FNO | Helped |
| increase training dataset size (30k to 100k trajectories) | SCENT · S5 dataset scalability | Rel-MSE | Helped |
Surprising findings
- FNO performs competitively on simulated challenging datasets despite lacking mesh independence.
- On AD-B, SCENT ranks second behind AROMA.
Limitations
- On AD-B, SCENT is outperformed by AROMA.
- FNO was modified to handle sparse data via zero-padding, which may affect its comparison.
- SCENT is the most computationally expensive per training step at NS-3 scale among the compared models.
Experimental setup
Datasets: Navier-Stokes NS-3, NS-4, NS-5 benchmarks simulated Navier-Stokes variants S1-S5 (new) AirDelhi AD-B, AD-T, AD-F
Benchmarks: Navier-Stokes NS-3 Navier-Stokes NS-4 Navier-Stokes NS-5 S1-S5 simulated datasets AirDelhi AD-B AirDelhi AD-T AirDelhi AD-F
Metrics: Rel-MSE MSE RMSE
Baselines: FNO OFormer CORAL AROMA DINO GNOT nearest-neighbor interpolation RBF interpolation IDW Random Forest XGBoost ARIMA
Builds on: CORAL OFormer Perceiver-style inducing point cross-attention Fourier features
Closest papers by meaning
- StreamPhy: Streaming Inference of High-Dimensional Physical Dynamics via State Space Models (NeurIPS 2026 · similarity 0.86)
- Wavelet-Driven Spatiotemporal Predictive Learning: Bridging Frequency and Time Variations (AAAI 2024 · similarity 0.85)
- Causal Deciphering and Inpainting in Spatio-Temporal Dynamics via Diffusion Model (NeurIPS 2024 · similarity 0.85)
- Arbitrarily Conditioned Hierarchical Flows for Spatiotemporal Events (NeurIPS 2026 · similarity 0.85)
- Multiple Physics Pretraining for Spatiotemporal Surrogate Models (NeurIPS 2024 · similarity 0.85)
- Learning Spatiotemporal Dynamical Systems from Point Process Observations (ICLR 2025 · similarity 0.85)
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.