Research map / Probabilistic and scientific modeling / Dynamical systems

Mechanism-Aware Ensemble Conditioning for Data-Limited Emulation of Extreme Events

NeurIPS 2026 · Probabilistic and scientific modeling › Dynamical systems · approach: Latent-space control and planning · 0 citations (OpenAlex)

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 a FiLM conditioning plug-in that feeds nudged-ensemble covariance statistics (a Jacobian-free proxy for finite-time instability geometry) into a residual-attention corrector for a 3D chaotic system and a STORN corrector for two-layer QG flow, improving data-limited extreme-event emulation.

Problem

Surrogates trained on short trajectories of chaotic systems learn bulk dynamics but miss the finite-time instability geometry that produces extreme events, hurting tail statistics and exceedance geometry.

Approach

Use the covariance of a small nudged stochastic ensemble around a coarse trajectory as a non-intrusive, Jacobian-free proxy for local finite-time deformation, and inject its compact statistics into an unchanged temporal backbone via FiLM.

Key results

• On QG, FiLM-STORN (T_train=50) reduces averaged exceedance-area error to 0.01047 vs 0.01574 for STORN T=50 and 0.01195 for STORN T=1000 (20x more data).
• On QG, FiLM-STORN reduces averaged exceedance-frequency error to 0.07199 vs 0.12118 for STORN T=50 and 0.11726 for STORN T=1000.
• In the 3D system, best FiLM setting reduces z3 E99 error by 26.3% (0.0056 to 0.0041) over the no-FiLM attention baseline; D_KL is worse for FiLM-STORN (0.0365 vs 0.0039 for STORN T=1000).

Surprising findings

• FiLM-STORN trained on 50 time units beats STORN trained on 1000 time units on high-threshold exceedance-area and frequency metrics, yet STORN T=1000 has much lower KL divergence and log-L1 error.
• Ensemble-derived log-L1 gain in the 3D system is only 2.5%, while the E99 tail gain is 26.3%.

Limitations

• Instability proxy depends on small-noise, local linearization and spectral-gap assumptions; noise level must be tuned to a narrow intermediate window.
• Ensemble simulation adds computational cost over a single coarse trajectory.
• Bulk statistics (KL, log-L1) can be under-corrected in severely data-limited settings; evaluation limited to one QG system and one backbone family plus a low-dimensional system.

Experimental setup

Datasets: two-layer quasi-geostrophic simulation (128x128 reference, 24x24 coarse) 3D controlled chaotic system trajectories

Benchmarks: two-layer QG held-out long trajectory 3D chaotic system test trajectory (T=500)

Metrics: KL divergence log-L1 distance exceedance frequency error exceedance area Wasserstein-1 distance 99th-percentile exceedance error

Baselines: STORN (T=50, no context) STORN (T=1000) CR (coarse-residual baseline) no-FiLM residual attention

Builds on: FiLM STORN optimally time-dependent modes nudging data assimilation Davis-Kahan theorem

Closest papers by meaning

• The Dynamic-Probabilistic Consistency Gap in Chaotic Surrogate Modeling (NeurIPS 2026 · similarity 0.88)
• First-Order Trajectory Matching: Fast Ensemble Predictions of Chaotic, Turbulent, Stochastic Systems (NeurIPS 2026 · similarity 0.87)
• Thermalizer: Stable autoregressive neural emulation of spatiotemporal chaos (ICML 2025 · similarity 0.85)
• Chaos Meets Attention: Transformers for Large-Scale Dynamical Prediction (ICML 2025 · similarity 0.84)
• Markovian Dynamics Enforcer: Feasibility Preserving Correction on Learned Dyanmics Manifolds (NeurIPS 2026 · similarity 0.84)
• Hierarchical Implicit Neural Emulators (NeurIPS 2025 · similarity 0.83)