[Paper Review] Physics-constrained deep learning postprocessing of temperature and humidity

Francesco Zanetta, Daniele Nerini+2 more|arXiv (Cornell University)|Dec 7, 2022

Model Reduction and Neural Networks4 citations

TL;DR

This paper proposes a physics-constrained deep learning framework for postprocessing surface temperature and humidity forecasts in Switzerland, enforcing thermodynamic equations directly in the neural network architecture. The method achieves physically consistent predictions without sacrificing performance, especially improving reliability under data scarcity, and ensures exact physical consistency by construction via differentiable analytic constraints.

ABSTRACT

Weather forecasting centers currently rely on statistical postprocessing methods to minimize forecast error. This improves skill but can lead to predictions that violate physical principles or disregard dependencies between variables, which can be problematic for downstream applications and for the trustworthiness of postprocessing models, especially when they are based on new machine learning approaches. Building on recent advances in physics-informed machine learning, we propose to achieve physical consistency in deep learning-based postprocessing models by integrating meteorological expertise in the form of analytic equations. Applied to the post-processing of surface weather in Switzerland, we find that constraining a neural network to enforce thermodynamic state equations yields physically-consistent predictions of temperature and humidity without compromising performance. Our approach is especially advantageous when data is scarce, and our findings suggest that incorporating domain expertise into postprocessing models allows to optimize weather forecast information while satisfying application-specific requirements.

Motivation & Objective

• To address physical inconsistencies in machine learning-based weather postprocessing, particularly violations of thermodynamic relationships between temperature and humidity.
• To improve forecast reliability and trustworthiness for downstream applications such as hydrology and renewable energy by enforcing physical laws in data-driven models.
• To evaluate whether integrating domain-specific analytic equations into deep learning enhances performance and robustness, especially when training data is limited.
• To explore whether physical constraints improve generalization to out-of-distribution conditions, such as extreme heat events.
• To demonstrate a simple, extensible framework for embedding scientific expertise into postprocessing models via differentiable equations.

Proposed method

• The method uses a fully connected neural network architecture where the final layer is constrained to satisfy the ideal gas law and Clausius-Clapeyron relationship between temperature and specific humidity.
• Physical consistency is enforced through a custom network layer that ensures the output satisfies the thermodynamic state equation T = f(q, p) by design, not via loss regularization.
• The model is trained using a standard mean absolute error (MAE) loss on observed temperature and specific humidity, with no additional physics-based loss terms.
• The approach is compared against unconstrained and loss-constrained deep learning baselines using the same training and evaluation protocol.
• The framework is evaluated under data scarcity by reducing training samples to 1% per station, and under out-of-distribution generalization by training on non-summer months and testing on summer.
• The method is implemented using differentiable operations, enabling end-to-end training while maintaining exact physical consistency.

Experimental results

Research questions

• RQ1 Can deep learning postprocessing models be made physically consistent by embedding analytic thermodynamic equations directly into the network architecture?
• RQ2 How does physical constraint affect model performance under data-scarce conditions, particularly when training samples are limited?
• RQ3 Does enforcing physical consistency improve generalization to extreme weather conditions not seen during training, such as summer heatwaves?
• RQ4 How does the architecture-constrained approach compare to unconstrained and loss-constrained models in terms of physical consistency and predictive accuracy?
• RQ5 Can this framework be extended to probabilistic forecasting and global-scale postprocessing in the future?

Key findings

• The architecture-constrained model enforces physical consistency to within machine precision, achieving exactly zero physical inconsistencies across all test cases.
• Under data scarcity (1% of training samples), the architecture-constrained model maintains physical consistency while showing only a minor performance degradation, unlike the unconstrained model, which exhibits increasing physical inconsistencies.
• The unconstrained model shows a significant rise in physical inconsistencies as training data is reduced, highlighting the value of physical constraints in low-data regimes.
• The model's predictive performance, measured by NMAE, remains comparable to unconstrained and loss-constrained models, with no degradation in accuracy despite the added physical constraints.
• Generalization to unseen summer temperature extremes was not significantly improved by physical constraints, indicating that such constraints help with interpolation but not with out-of-distribution extrapolation.
• The method is extensible and can be applied to any differentiable physical equation, enabling integration of domain expertise into postprocessing models in a modular and principled way.