Discovering Equations for the Thermodynamics of South Asian Monsoon Precipitation

Savannah L. Ferretti, Jane W. Baldwin, Michael S. Pritchard, Tom BeuclerarXiv 2026 · arXiv

TL;DR

Researchers have used "symbolic distillation"—a hybrid of neural networks and symbolic regression—to discover simple, interpretable equations for South Asian monsoon precipitation. These data-driven equations outperform traditional physics-based diagnostics by 50% while using 1,000x fewer parameters than deep learning models, revealing that moisture and instability act as competing, rather than additive, controls on rainfall.

What Problem Does This Paper Address?

Predicting monsoon rainfall is a high-stakes challenge for South Asia, yet the underlying thermodynamics remain surprisingly debated. We know rain depends on the vertical structure of the atmosphere (temperature and moisture), but this dependence is "nonlocal"—what happens at 2km up affects what happens at the surface.

Current approaches fall into two camps:

  1. Process-oriented diagnostics: Simple metrics like Lower-Tropospheric Buoyancy ($B_L$) that rely on rigid, pre-defined assumptions about atmospheric layers.
  2. Deep Learning: Highly accurate "black box" models that are nearly impossible to interpret physically.

The authors sought a "middle way": can we use AI to discover the underlying mathematical laws of precipitation that are both accurate and physically transparent?

Methods

The authors employed a two-stage pipeline called Symbolic Distillation:

  1. Integration Kernel Learning (IKL): Instead of assuming which atmospheric layers matter, they trained a neural network to learn "kernels"—weighting functions that compress entire vertical profiles (Relative Humidity, Equivalent Potential Temperature) into single scalar features. This effectively performs data-driven dimensionality reduction on the vertical column.
  2. Symbolic Regression (SR): Using the PySR package, the researchers searched for the simplest closed-form equations that link these learned features and surface heat fluxes to precipitation.
  3. Physical Constraint Enforcement: To ensure the AI didn't "cheat" with unphysical math, they tested the equations against three Physical Constraints (PCs), such as ensuring that increasing moisture leads to more (not less) rain. They even manually refined one model (SR-ALL-PC) to guarantee physical consistency across all inputs.

Datasets

The study utilized 21 years (2000–2020) of ERA5 reanalysis data during the peak monsoon months (June–August).

  • Target: Total precipitation.
  • Predictors: Vertical profiles of Relative Humidity (RH), equivalent potential temperature ($\theta_e$), and saturated equivalent potential temperature ($\theta_e^*$), plus surface variables like Land Fraction (LF) and Latent/Sensible Heat Fluxes (LHF/SHF).
  • Scale: Data was interpolated to a 1.0° grid and 3-hour windows, totaling approximately 10 million samples.

Key Insights and Ideas

  • Competition Over Addition: Traditional metrics often add moisture and instability terms together. The discovered equations suggest a competitive structure using a max(M, I) operator. Rain intensifies based on whichever "pathway"—moisture or instability—is more dominant, rather than a simple sum of the two.
  • The "Distinguisher" for Extremes: While instability determines if it will rain, lower-tropospheric moisture is what distinguishes moderate rain from heavy, organized precipitation states.
  • Oceanic Coupling: Over the ocean, sensible heat flux ($SHF$) isn't an independent driver. Its impact on rain is strongly coupled to near-surface thermodynamics; it improves predictions most when the air is already warm and moist.
  • Vertical Sensitivity: The learned kernels revealed that moisture matters most at the top of the lower troposphere ($\sim$500 hPa), while parcel energy ($\theta_e$) is most critical right at the surface.

Key Results

  • Performance vs. Complexity: The best symbolic equation (SR-ALL-PC) achieved an $R^2 \approx 0.42$. While lower than the full "black box" neural network ($R^2 \approx 0.58$), it vastly outperformed the standard $B_L$ buoyancy diagnostic ($R^2 \approx 0.27$).
  • Efficiency: The symbolic models recover about 85% of the neural network's skill while using roughly three orders of magnitude fewer parameters.
  • Diurnal Improvement: Including surface heat fluxes significantly improved the prediction of the oceanic diurnal cycle, reaching 74% of the observed range (compared to 42% for atmosphere-only models).

Limitations and Future Work

  • The Land Diurnal Cycle: The equations still struggle to capture the timing of afternoon convection over land, likely because the current model only looks at the concurrent state and ignores the "memory" of the morning's heating.
  • Causality: Because the predictors and targets are concurrent, the equations describe correlations. It remains unclear if surface fluxes cause rain or if the rain (via cold pools) causes the flux changes.
  • Missing Dynamics: The study focused purely on thermodynamics. The authors suggest that adding "dynamical" features—like low-level wind convergence or interactions with South Asia’s complex topography—could account for the remaining unexplained variability.

Summary updated

Abstract

South Asian monsoon precipitation depends on the thermodynamic structure of the lower troposphere, but how its ingredients combine to shape precipitation remains unclear, partly because this dependence is nonlocal in the vertical. We use symbolic distillation, which combines neural networks and symbolic regression, to discover interpretable precipitation equations from 21 years of June-August reanalysis data. Neural networks first learn the vertical thermodynamic structure most relevant to precipitation, condensing each variable into a single interpretable feature not tied to a particular vertical grid. Then, symbolic regression derives closed-form equations from these features and the surface state. Our best equation explains 1.5 times as much precipitation variance as a widely used buoyancy-based diagnostic and recovers about 85% of neural network skill while using about a thousand times fewer parameters. Unlike the buoyancy diagnostic, which adds moisture and instability, the equations treat them as competing controls, with the heaviest precipitation occurring when the lower free troposphere is humid. Over ocean, sensible heating is associated with additional precipitation mainly when near-surface air is warm and moist, improving predicted precipitation and its diurnal cycle. Although the learned features were optimized only for precipitation prediction, the equations largely respond to each feature as expected from convective physics, consistent with the reanalysis. These responses reduce the dependence of monsoon precipitation on the entire vertical column to a few simple, testable constraints that could guide data-driven precipitation models to make them more physically consistent.
BibTeX
@misc{ferretti2026discovering,
  author = {Savannah L. Ferretti and Jane W. Baldwin and Michael S. Pritchard and Tom Beucler},
  title = {{Discovering Equations for the Thermodynamics of South Asian Monsoon Precipitation}},
  year = {2026},
  eprint = {2611.14401},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.14401},
}
Figure