Wang et al. (2025) Causal machine learning uncovers conditions for convective intensification driven by organic and sulfate aerosols

Identification

• Journal: Scientific Reports
• Year: 2025
• Date: 2025-12-29
• Authors: Die Wang, LI Jie-xi, Jun Lu
• DOI: 10.1038/s41598-025-28939-x

Research Groups

• Environmental Science and Technologies Department, Brookhaven National Laboratory, Upton, NY, USA
• Institute for Atmospheric and Climate Science, ETH Zurich, Zurich, Switzerland
• Applied Mathematics and Statistics, Stony Brook University, Stony Brook, NY, USA
• School of Public Health, University of Illinois Chicago, Chicago, IL, USA

Short Summary

This study applies a novel causal machine learning framework to high-resolution observations near Houston, TX, to investigate the causal links between organic and sulfate aerosols and deep convective clouds (DCCs). It finds that a direct causal link from aerosols to DCCs is uncommon (less than 35% of scenarios) but, when present, can substantially enhance DCC core heights by approximately 1.7 kilometers, particularly in warmer cloud regions and under sea breeze conditions.

Objective

• To provide observational evidence for or against aerosol impacts on deep convective clouds (DCCs) in a coastal urban environment (Houston, TX).
• To clarify the magnitude, sign, and meteorological dependence of aerosol effects on DCCs, specifically uncovering conditions for convective intensification driven by organic and sulfate aerosols.
• To apply a statistically rigorous causal discovery-inference pipeline to isolate and quantify causal effects from observational datasets in the field of aerosol-cloud interaction.

Study Configuration

• Spatial Scale: Observations near Houston, TX, USA. Radar data covered a 400 km × 400 km domain. Deep convective clouds (DCCs) were analyzed within radii of 30 km, 40 km, and 50 km from the ARM main site.
• Temporal Scale: Data collected from June to September 2022, as part of the TRacking Aerosol Convection Interactions ExpeRiment (TRACER). Aerosol properties were averaged over a 1-hour window centered on sounding launch times, and DCC echo top heights (ETHs) were observed within 6 hours after launch. Radiosondes were launched 4–7 times per day.

Methodology and Data

• Models used:
  • Causal Discovery: Causal Additive Model with Unobserved Variables (CAM-UV), nonlinear Non-combinatorial Optimization via Trace Exponential and Augmented Lagrangian for Structure Learning (NOTEARS).
  • Causal Inference: Double/Debiased Machine Learning (DML) with Random Forest, Gradient Boosting, or Multiple Linear Regression as base-learner models.
  • Convective Cell Tracking: Lagrangian cell tracking algorithm.
  • Synoptic Regime Classification: Self-organizing map (SOM).
• Data sources:
  • High-resolution observations from the Department of Energy’s Atmospheric Radiation Measurement (ARM) TRACER campaign.
  • NOAA S-band Doppler radar KHGX-Houston (Level-II reflectivity data).
  • ARM balloon-borne sounding system (SONDE) observations for pre-convective meteorological conditions (CAPE, LCL, LNB, ELR, CIN, LFC, WSR, RH).
  • ARM Aerosol Observing System (AOS) surface-based measurements for aerosol conditions (CCN number concentrations at 0.1%, 0.2%, 0.4%, 0.6%, 0.8%, 1% supersaturation; total aerosol number concentrations (N{cn}) (10–3000 nm) and (N{ufp}) (3–3000 nm)).
  • ERA5 global reanalysis for synoptic regime classification.

Main Results

• A direct causal link from aerosols (ARO) to deep convective cloud (DCC) intensity (Echo Top Height, ETH) was identified in 10% (CAM-UV) to 35% (NOTEARS) of scenarios, with an overall validation rate of 27% across all sensitivity tests, indicating a conditional and nonlinear relationship.
• When aerosol impacts were detected, they could be substantial, enhancing DCC core heights by approximately 1.7 km on average.
• 92% of the aerosol-induced invigoration effect was concentrated in warmer-phase cloud regions (using 30-dBZ ETH as a proxy), compared to 57% when using the 15-dBZ ETH proxy.
• The presence of sea breezes significantly increased the proportion of valid cases (31%) and invigoration effects (95%), with an estimated average enhancement of 1.9 km, which is 0.8 km greater than the estimate from the full DCC sample.
• Using total aerosol number concentrations ((N{cn}) and (N{ufp})) as exposure variables resulted in nearly all estimated effects being positive (>99% invigoration) and the highest fraction of aerosol-sensitive scenarios.
• The choice of causal graph assumption was critical for inference, with CAM-UV (one adjustable covariate, LCL) yielding a higher validation rate (31%) compared to NOTEARS (three adjustable covariates, ELR, CAPE, LNB) (22%).
• The estimated aerosol effect and associated uncertainty decreased as a larger radius from the ARM site was considered, with results stabilizing beyond 40 km. A radius of approximately 30 km was suggested as the effective range of DCC–aerosol interactions.

Contributions

• This study represents one of the first applications of a full causal discovery–inference pipeline (combining CAM-UV/NOTEARS and Double/Debiased Machine Learning) to the field of aerosol–cloud interaction using solely observational datasets.
• It offers a statistically rigorous and generalizable framework for isolating and estimating causal effects, moving beyond traditional correlation-based analyses in atmospheric science.
• Provides crucial observational evidence for the conditional, nonlinear, and context-dependent nature of aerosol impacts on DCCs, helping to resolve long-standing debates in the atmospheric science community.
• Highlights the critical importance of explicit causal structure specification in guiding robust inference and mitigating biases from unaddressed confounders.
• Demonstrates the resilience of the DML framework to biases introduced by specific base-learner models, enhancing confidence in the results.
• Offers valuable constraints for the development and improvement of next-generation convection-permitting models and informs the design of physics-aware parameterizations.

Funding

• U.S. DOE Early Career Research Program
• Atmospheric System Research (ASR) program
• Office of Workforce Development for Teachers and Scientists (WDTS) under the Science Undergraduate Laboratory Internships Program (SULI)
• Brookhaven Science Associates, LLC, under Contract DE-SC0012704
• Swiss Federal Institute of Technology Zurich (Open access funding)

Citation

@article{Wang2025Causal,
  author = {Wang, Die and Jie-xi, LI and Lu, Jun},
  title = {Causal machine learning uncovers conditions for convective intensification driven by organic and sulfate aerosols},
  journal = {Scientific Reports},
  year = {2025},
  doi = {10.1038/s41598-025-28939-x},
  url = {https://doi.org/10.1038/s41598-025-28939-x}
}