Introduces the Perception-Physics Paradox and TC-Bench benchmark demonstrating that vision foundation models rely on visual shortcuts that fail in intense regimes rather than achieving scientific alignment via structural isomorphism.
Canonical reference
Salva Rühling Cachay, Miika Aittala, Hailey James, and Rose Yu
Canonical reference. 80% of citing Pith papers cite this work as background.
22
Pith papers citing it
2
external citations · external index
Background
80% of classified citations
citation-role summary
background 5
citation-polarity summary
years
2026 22roles
background 5representative citing papers
AI weather models may simulate the atmosphere via particle positions in latent space whose updates follow gradient flow on a learned free energy functional rather than conventional physical equations.
A single neural operator can approximate the map from arbitrary joint densities to their conditionals, backed by new continuity results and illustrated on Gaussian mixtures.
Cast3 translates NWP principles into a data-driven model using cubed-sphere grids, super-ensembles, and generative nudging to achieve state-of-the-art ensemble predictions that outperform baselines.
Historically trained ML weather emulators quantify fast precipitation changes from CO2 perturbations and produce results that agree with Earth System Models.
SnareNet introduces a repair layer that navigates the range space of constraints plus adaptive relaxation training to enforce hard non-convex constraints on neural network outputs more reliably than prior methods.
Augments the energy score objective for sample-based generative models with a differentiable decision loss that is itself a proper scoring rule, yielding targeted improvements on cost-sensitive regions in synthetic and real tasks.
Rank-1 lattice points and hyperbolic crosses are shown to improve the generalization error and efficiency of Fourier neural operators for PDE approximation on the torus.
PredHydro-Net is a new dual-decoding architecture with wavelet spectral matching and adversarial training that outperforms Earthformer, PredRNNv2, and GFS on extreme-event detection and spectral fidelity in 72-hour global hydrometeor forecasts.
Training climate emulators on random-CO2 runs that break SST–CO2 correlation, plus an energy constraint, yields a data-efficient model that works on AMIP+4K and abrupt 4xCO2 cases prior ACE models mishandled.
Samudra 2 scales autoregressive neural ocean emulators to finer resolutions with architectural tweaks and dynamic loss, raising upper-ocean temperature R² from 0.56 to 0.87 at 1° and recovering mesoscale features.
Njord introduces a probabilistic GNN model using latent variables and adaptive K-means meshes for ensemble ocean forecasting with uncertainty estimates on global and regional domains.
ShardTensor is a domain-parallelism system for SciML that enables flexible scaling of extreme-resolution spatial datasets by removing the constraint of batch size one per device.
ESFM is a single open foundation model that unifies heterogeneous Earth data sources and forecasts missing regions while preserving inter-variable physical relationships.
A discrete-time energy-conserving neural map with FDT-derived causal regularization reproduces stationary statistics and forced responses of CdV and Lorenz-96 turbulence from unperturbed data alone.
Multi-scale wavelet transformers learn operator dynamics of chaotic systems in the wavelet domain, yielding lower errors and higher spectral fidelity on benchmarks and ERA5 climate data.
HealDA supplies ML-based initial conditions for AI weather models that produce forecasts trailing ERA5-initialized runs by less than one day of effective lead time, with the skill gap arising mainly from initial error size.
LSTM and 1D CNN emulators replicate a 1D marine biogeochemistry model at daily resolution, remain stable over decades, reproduce spring bloom timing years ahead, and outperform the parent model on reanalysis-driven forecasts for key variables.
Otter Weather is a spatiotemporal model that outperforms NWP baselines by 9.6% at 24h lead with under 3.5 A100-days training and extends efficiency gains to probabilistic forecasting via CRPS.
CRPS-trained ensembles achieve better uncertainty reliability and speed than latent generative models for probabilistic emulation of 2D physical systems.
A standard U-Net with MAE pre-training plus short CRPS fine-tuning and MC Dropout matches GenCast and IFS ENS probabilistic skill at 1.5° while cutting training and inference cost by over 10×.
Scaling laws for weather models exhibit strong cross-channel and cross-horizon heterogeneity, where globally pooled metrics appear favorable while many individual channels degrade at longer leads.
citing papers explorer
-
The Perception-Physics Paradox: Probing Scientific Alignment with TC-Bench
Introduces the Perception-Physics Paradox and TC-Bench benchmark demonstrating that vision foundation models rely on visual shortcuts that fail in intense regimes rather than achieving scientific alignment via structural isomorphism.
-
The physics of AI weather models
AI weather models may simulate the atmosphere via particle positions in latent space whose updates follow gradient flow on a learned free energy functional rather than conventional physical equations.
-
One Operator for Many Densities: Amortized Approximation of Conditioning by Neural Operators
A single neural operator can approximate the map from arbitrary joint densities to their conditionals, backed by new continuity results and illustrated on Gaussian mixtures.
-
Cast3: Translating numerical weather prediction principles into data-driven forecasting
Cast3 translates NWP principles into a data-driven model using cubed-sphere grids, super-ensembles, and generative nudging to achieve state-of-the-art ensemble predictions that outperform baselines.
-
Examining Fast Radiatively Driven Responses Using Machine-Learning Weather Emulators
Historically trained ML weather emulators quantify fast precipitation changes from CO2 perturbations and produce results that agree with Earth System Models.
-
SnareNet: Flexible Repair Layers for Neural Networks with Hard Constraints
SnareNet introduces a repair layer that navigates the range space of constraints plus adaptive relaxation training to enforce hard non-convex constraints on neural network outputs more reliably than prior methods.
-
Decision-Aware Training for Sample-Based Generative Models
Augments the energy score objective for sample-based generative models with a differentiable decision loss that is itself a proper scoring rule, yielding targeted improvements on cost-sensitive regions in synthetic and real tasks.
-
Fourier Neural Operators with rank-1 lattice points and hyperbolic cross
Rank-1 lattice points and hyperbolic crosses are shown to improve the generalization error and efficiency of Fourier neural operators for PDE approximation on the torus.
-
Physics-Guided Dual Decoding and Spectral Supervision for Global 3D Hydrometeor Prediction
PredHydro-Net is a new dual-decoding architecture with wavelet spectral matching and adversarial training that outperforms Earthformer, PredRNNv2, and GFS on extreme-event detection and spectral fidelity in 72-hour global hydrometeor forecasts.
-
Disentangling the effects of sea surface temperature and CO$_2$ in global machine learned weather-climate emulators
Training climate emulators on random-CO2 runs that break SST–CO2 correlation, plus an energy constraint, yields a data-efficient model that works on AMIP+4K and abrupt 4xCO2 cases prior ACE models mishandled.
-
Samudra 2: Scaling Ocean Emulators across Resolutions
Samudra 2 scales autoregressive neural ocean emulators to finer resolutions with architectural tweaks and dynamic loss, raising upper-ocean temperature R² from 0.56 to 0.87 at 1° and recovering mesoscale features.
-
Njord: A Probabilistic Graph Neural Network for Ensemble Ocean Forecasting
Njord introduces a probabilistic GNN model using latent variables and adaptive K-means meshes for ensemble ocean forecasting with uncertainty estimates on global and regional domains.
-
ShardTensor: Domain Parallelism for Scientific Machine Learning
ShardTensor is a domain-parallelism system for SciML that enables flexible scaling of extreme-resolution spatial datasets by removing the constraint of batch size one per device.
-
Earth System Foundation Model (ESFM): A unified framework for heterogeneous data integration and forecasting
ESFM is a single open foundation model that unifies heterogeneous Earth data sources and forecasts missing regions while preserving inter-variable physical relationships.
-
Physics constraints and response validation in discrete-time reduced-order modeling: from idealized turbulent systems to climate dynamics
A discrete-time energy-conserving neural map with FDT-derived causal regularization reproduces stationary statistics and forced responses of CdV and Lorenz-96 turbulence from unperturbed data alone.
-
Multi-Scale Wavelet Transformers for Operator Learning of Dynamical Systems
Multi-scale wavelet transformers learn operator dynamics of chaotic systems in the wavelet domain, yielding lower errors and higher spectral fidelity on benchmarks and ERA5 climate data.
-
HealDA: Highlighting the importance of initial errors in end-to-end AI weather forecasts
HealDA supplies ML-based initial conditions for AI weather models that produce forecasts trailing ERA5-initialized runs by less than one day of effective lead time, with the skill gap arising mainly from initial error size.
-
Deep learning model emulators for marine biogeochemistry forecasting from days to decades
LSTM and 1D CNN emulators replicate a 1D marine biogeochemistry model at daily resolution, remain stable over decades, reproduce spring bloom timing years ahead, and outperform the parent model on reanalysis-driven forecasts for key variables.
-
Otter Weather: Skillful and Computationally Efficient Medium-Range Weather Forecasting
Otter Weather is a spatiotemporal model that outperforms NWP baselines by 9.6% at 24h lead with under 3.5 A100-days training and extends efficiency gains to probabilistic forecasting via CRPS.
-
Reliability of Probabilistic Emulation of Physical Systems
CRPS-trained ensembles achieve better uncertainty reliability and speed than latent generative models for probabilistic emulation of 2D physical systems.
-
U-Cast: A Surprisingly Simple and Efficient Frontier Probabilistic AI Weather Forecaster
A standard U-Net with MAE pre-training plus short CRPS fine-tuning and MC Dropout matches GenCast and IFS ENS probabilistic skill at 1.5° while cutting training and inference cost by over 10×.
-
Towards Scaling Law Analysis For Spatiotemporal Weather Data
Scaling laws for weather models exhibit strong cross-channel and cross-horizon heterogeneity, where globally pooled metrics appear favorable while many individual channels degrade at longer leads.