The instrument at the heart of the effort is the Ka-band cloud radar, which transmits millimeter-wavelength pulses exquisitely sensitive to the tiny droplets and ice crystals that make up non-precipitating and weakly precipitating clouds. Unlike conventional weather radars tuned to rain, these systems can peer through thick, multi-layered cloud decks and return continuous vertical profiles up to 18 kilometers above the ground. Of roughly 120 radars operating nationwide in 2024, the researchers retained 80 stations that logged at least 100 days of valid observations each, a quality threshold that made a robust national analysis possible for the first time. Before 2024, fewer than 30 such radars were in operation, far too sparse for a coherent countrywide picture.

Raw radar echoes, however, are notoriously contaminated. Near-surface dust, insect swarms, turbulence, and instrumental noise can masquerade as cloud, while a radar’s sensitivity to faint high clouds fades with distance. The team addressed both problems with an improved retrieval algorithm that replaces the traditional fixed reflectivity cutoff with a height-dependent noise-equivalent threshold, so that weak echoes aloft are judged against a realistic sensitivity curve rather than a single number. A second, variance-based filter exploits the fact that true cloud layers are temporally coherent, whereas clutter flickers erratically; a three-minute sliding window separates the two. Validation against collocated radiosondes showed discrepancies of well under half a kilometer for thin low and mid-level clouds and no more than about a kilometer for deep precipitating systems.

The headline numbers are striking. Across the network, clouds were present 56.7 percent of the time on average, and single-layer clouds dominated, accounting for 34.7 percent of all observations, roughly 62 percent of cloudy moments. Two-layer clouds occurred 14.7 percent of the time and multi-layer systems, with three to five distinct decks, just 7.1 percent. Geography mattered enormously. Single-layer clouds prevailed over the arid basins and plateaus of northwestern China, where dry, stable air suppresses vertical motion, while multi-layer clouds clustered over the humid southeast, where monsoon moisture and strong solar heating fuel deep, vertically complex convection. The highest overall cloud frequencies appeared over the southeastern Tibetan Plateau and the Sichuan Basin, where topographic lifting and moisture convergence conspire to fill the sky.

Cloud base height told a vivid seasonal story. Nationally averaged bases climbed from 2.15 kilometers in winter to 3.24 kilometers in spring, 3.95 kilometers in summer, and back down to 2.78 kilometers in autumn, a progression the authors attribute to the interplay of surface heating, boundary-layer depth, and the lifting condensation level. Counterintuitively, cloud bases over the Tibetan Plateau sat distinctly lower than over surrounding lowlands during spring and summer, because the Plateau’s cold ambient air lowers the condensation level for a given humidity even as intense surface heating deepens mixing. Cloud tops followed a similar seasonal rhythm, averaging 7.51 kilometers in summer against 4.70 kilometers in winter, and vertically the distribution of cloud tops was bimodal, with one population peaking between 1 and 3 kilometers and a second between 6 and 12 kilometers, the signature of coexisting shallow stratiform and deep convective regimes.

The radar network’s minute-by-minute sampling also exposed a diurnal drama that twice-daily radiosondes have long missed. In summer, cloud occurrence between 3 and 9 kilometers rose from an early-morning minimum to a pronounced late-afternoon and evening peak between 16:00 and 20:00 local time, as solar heating destabilized the boundary layer and ignited moist convection that punched into the upper troposphere. Winter flipped the script: clouds stayed largely below 3 kilometers and peaked near sunrise, when minimum temperatures under a stable nocturnal boundary layer pushed relative humidity high enough to condense shallow stratiform decks. Spring and autumn traced transitional patterns in between, with afternoon mid-tropospheric peaks in spring and increasingly even, low-topped cloudiness in autumn.

What drives all this variability? The team matched the radar retrievals hourly against ERA5 reanalysis fields and wind-profiler measurements to isolate three candidate controls: near-surface specific humidity, lower-tropospheric stability, and vertical wind shear. Moisture emerged as the dominant driver. Specific humidity correlated positively with both cloud base and cloud top at more than 93 percent of stations, and with cloud thickness at 71 percent, confirming that abundant low-level moisture sustains the buoyancy needed for deep vertical development. Stability acted as a powerful brake: a terrain-following measure of lower-tropospheric stability correlated negatively with cloud base at 97.5 percent of stations and with cloud top at 93.8 percent, meaning strongly stratified air confines clouds to shallow, low-based layers regardless of how moist the surface is.

Wind shear played a subtler, conditional role. Strong vertical shear correlated negatively with cloud base, top, and thickness at roughly three-quarters of stations, consistent with shear-enhanced entrainment of dry air that erodes buoyancy and shreds organized updrafts. Yet about a quarter of stations, mainly in the northwest, showed the opposite sign, echoing the well-known optimal shear framework in which moderate shear organizes convection rather than destroying it. Joint analyses made the hierarchy explicit: the tallest, thickest clouds required both high humidity and low stability, while under stable conditions cloud bases stayed low no matter the moisture or shear. Moisture was the engine; stability and shear were the governors.

Perhaps the most novel finding concerns the ground itself. Comparing cloud base heights across land-cover types mapped by MODIS, the researchers found that low clouds rooted lowest over densely vegetated surfaces, where evapotranspiration moistens the boundary layer and drops the condensation level, and highest over barren land, where moisture scarcity lifts it. The contrast sharpened dramatically in the afternoon. Single-layer low clouds observed between 14:00 and 18:00 local time had bases averaging 21.7 percent higher than the all-hour climatology, with the enhancement reaching 41.7 percent over evergreen broadleaf forests and 33.3 percent over croplands, but only a few percent over water bodies whose thermal inertia damps the diurnal boundary-layer cycle. The result is a crisp observational fingerprint of land-atmosphere coupling, showing that what grows on the surface measurably shapes where clouds form above it.

For modelers, the implications are immediate. The observed sensitivities of cloud base and top to humidity and stability provide quantitative benchmarks for calibrating the boundary-layer and convective parameterizations that dominate cloud-related uncertainty in weather and climate prediction, while the land-cover contrasts offer a test of how faithfully land surface models capture surface-atmosphere feedbacks. The authors caution that a single year of data cannot establish trends, and they plan to extend the analysis to multi-year records to probe interannual variability and the role of cloud structure in extreme weather. But as a first national snapshot, the study demonstrates what a coordinated ground-based radar network can do: turn one of climate science’s murkiest uncertainties into a measured, mappable, and model-testable quantity, one kilometer of sky at a time.

Subject of Research: Cloud vertical structure across China observed by a national Ka-band cloud radar network and its thermodynamic, dynamical, and land-surface controls

Article Title: Cloud vertical structure across China from a national Ka-band cloud radar network: thermodynamic, dynamical, and land-surface controls

Article References: Xu, H., Guo, J., Deng, J., Yang, R., Meng, D., Zhang, Z., Li, N., Sun, Y., Jiang, S., Chen, T., Chen, J., Zeng, L., Zhou, Y., & Tong, B. (2026). Cloud vertical structure across China from a national Ka-band cloud radar network: thermodynamic, dynamical, and land-surface controls. Atmospheric Chemistry and Physics, 26(19), 14303-14327. https://doi.org/10.5194/acp-26-14303-2026

Image Credits: AI Generated

DOI: 10.5194/acp-26-14303-2026

Keywords: cloud vertical structure, Ka-band cloud radar, cloud base height, cloud top height, lower-tropospheric stability, vertical wind shear, specific humidity, land-atmosphere coupling, Tibetan Plateau, East Asian monsoon, cloud parameterization, China Meteorological Administration

The stakes are far higher than one snowy field site might suggest. Mixed-phase clouds, in which supercooled liquid droplets and ice crystals coexist between 0 and minus 40 degrees Celsius, dominate the mid- and high-latitudes and exert a powerful influence on Earth’s radiative balance and precipitation. At high latitudes, where ice-albedo feedbacks and polar warming amplification are strongest, these clouds play an outsized role in the climate system. Yet their behavior hinges on ice-nucleating particles, or INPs, which are astonishingly scarce: for much of the heterogeneous freezing spectrum, they represent as little as one in a million of all atmospheric particles. Predicting their occurrence and abundance remains one of the great uncertainties in climate modeling, which is why the Hyytiälä team turned to an unconventional ally in their quest.

Between February and June 2018, the researchers deployed two continuous-flow diffusion chambers, the Portable Ice Nucleation Chamber known as PINC and its upgraded successor PINCii, to sample ambient air with unprecedented temporal resolution. Each instrument operated in a repeating cycle of five minutes of particle-free background measurement followed by fifteen minutes of ambient sampling, yielding one INP concentration data point every twenty minutes. PINC ran from 19 February to 2 April at a lamina temperature of minus 31 degrees Celsius, while PINCii took over from 22 April to 10 June at minus 32 degrees, both at a relative humidity with respect to water of 105 percent. The two chambers agreed within a factor of two during overlapping checks, and their sequential operation provided near-continuous high-frequency coverage across the dramatic seasonal transition from snow-covered winter to the awakening boreal spring.

The real innovation lay in pairing these measurements with the staggering observational infrastructure of SMEAR II, a station embedded in the Integrated Carbon Observation System, the Aerosol, Clouds and Trace Gases Research Infrastructure, and the Swedish Infrastructure for Ecosystem Science. The team interrogated 509 individually monitored variables spanning meteorology, radiation, soil conditions, aerosol physics, and trace gas chemistry. After quality screening that removed variables plagued by missing values, constant readings, and redundant duplicates, 84 parameters survived for analysis. All were harmonized to a common twenty-minute time base, creating a multidimensional dataset against which the INP measurements could be tested for hidden structure.

The machine-learning toolkit included random forest and decision tree models for feature importance ranking, alongside principal component analysis and K-means clustering for dimensionality reduction and pattern exploration. Crucially, the authors emphasize that these were exploratory tools for hypothesis generation rather than predictive modeling; no train-test splits or cross-validation were performed. To guard against statistical flukes, the random forest analysis was repeated under 50 independent random seeds, and the top-ranked variables remained stable across all of them. When many variables are screened simultaneously, the probability of finding spurious correlations rises sharply, and the researchers are explicit that their identified associations require independent validation before any causal claims can be made.

The rankings that emerged were revealing. The six highest-ranked predictors included nitrate aerosol mass measured by an aerosol mass spectrometer, the number concentration of particles between 250 and 1000 nanometers, fluorescent biological aerosol particle counts from a Wideband Integrated Bioaerosol Sensor, gas-phase acetone concentrations, cloud condensation nuclei at low supersaturation, and the time air masses had spent over land according to back-trajectory analysis. Fluorescent particles, detected through dual ultraviolet excitation, serve as tracers for primary biological aerosol, a well-studied INP source in forest environments. Surprisingly, indicators of black carbon also ranked highly, despite laboratory evidence that fresh soot is a poor ice nucleator, though atmospheric ageing, oxidation, and organic coating may enhance its ice activity, and a previous study at the same site independently reported short-timescale correlations between black carbon and INPs.

The seasonal split proved to be the study’s most striking finding. The winter measurements from PINC showed no meaningful correlation with any of the 84 examined variables; the fitted exponents in simple power-law relationships collapsed to near zero, meaning predicted concentrations were essentially flat regardless of what the environment was doing. In contrast, the spring and summer data from PINCii responded to ambient aerosol properties, with fluorescent particle concentration and nitrate mass both yielding moderate predictive skill, achieving adjusted R-squared values of roughly 0.5. The authors suggest that a local biogenic contribution, swelling as the ecosystem emerged from winter, augments a transported background aerosol during the growing season, making local proxies genuinely informative only in that regime.

Equally telling was the failure of established parameterizations. Widely used formulations from DeMott and colleagues and from Tobo and colleagues, which predict INP concentrations from temperature and the abundance of particles larger than half a micron, produced negative coefficients of determination against the Hyytiälä data, meaning they performed worse than simply guessing the observational mean. The Tobo parameterization, calibrated in a temperate North American forest where large particles are predominantly primary biological aerosol, systematically overpredicted INPs at the boreal site, partly because frequent new particle formation events there grow secondary organic aerosol into the relevant size range without adding ice-active particles. The team also found the INP concentration distributions to be broadly log-normal, a signature of random dilution and mixing that points to long-range transport from diverse, distant sources as the dominant winter supply mechanism, a conclusion echoing earlier work at the same station.

The deeper lesson is philosophical as much as technical. Correlation, even when machine-assisted and replicated across dozens of random seeds, does not illuminate causation, particularly when the target of interest is a vanishingly small fraction of the total aerosol and its identity remains undetermined without single-particle analysis of ice crystal residuals. The authors argue that the very question of why local measurements cannot predict INPs is itself scientifically valuable, because it constrains the problem space and redirects the field toward approaches that explicitly account for air mass history. They propose that future campaigns pair high-frequency INP measurements with source-apportionment tools such as back-trajectory analysis, receptor modeling, and, where possible, single-particle ice residual analysis, and point to efforts like the ACTRIS cloud in-situ initiative as platforms for such integrated studies.

For climate modelers, the message is clear: site-specific parameterizations rooted in local biogenic and chemical proxies, such as fluorescent particle concentrations and nitrate aerosol mass, offer the most promising path forward for boreal environments during the growing season, while winter INP populations in southern Finland appear governed by processes far beyond the forest. As the Arctic warms and mixed-phase clouds shift the energy balance of the high latitudes, understanding where ice-forming particles come from, and accepting that some of the answers lie hundreds or thousands of kilometers upwind, may prove essential to forecasting the future of the frozen North.

Subject of Research: Ice-nucleating particle concentrations and machine-learning variable screening in a boreal forest environment

Article Title: Exploring ice nucleation particle concentrations in a boreal environment: limits of machine-learning-assisted variable screening

Article References: Wu, Y., Brasseur, Z., Castarède, D., Heikkilä, P., Keskinen, J., Möhler, O., Kulmala, M., Petäjä, T., Thomson, E. S., & Duplissy, J. (2026). Exploring ice nucleation particle concentrations in a boreal environment: limits of machine-learning-assisted variable screening. Aerosol Research, 4(2), 293-309. https://doi.org/10.5194/ar-4-293-2026

Image Credits: AI Generated

DOI: 10.5194/ar-4-293-2026

Keywords: ice-nucleating particles, mixed-phase clouds, boreal forest, SMEAR II, machine learning, random forest, aerosol chemistry, fluorescent bioaerosol, long-range transport, cloud parameterization, HyICE-2018, climate modeling

The researchers ran a set of sensitivity experiments with RegCM4 version 4.5.0, driving the model with ERA-Interim reanalysis data at the lateral boundaries and NOAA Optimum Interpolation weekly sea surface temperatures at the lower boundary. They then evaluated the simulations against the ERA5 reanalysis for precipitation, two-meter air temperature, atmospheric circulation, and water vapor flux divergence, and against the CLARA-A2 satellite cloud products from the CM SAF archive for cloud properties. The baseline control experiment, labeled CTRL, used the model’s default configuration, while three alternative experiments paired different cloud microphysics schemes with different cloud fraction diagnostic schemes, including the Nogherotto–Tompkins multi-phase microphysics scheme and the Xu–Randall semiempirical cloud fraction formulation.

The verdict on the default configuration was unflattering. The CTRL experiment clearly overestimated both summer precipitation and near-surface air temperature over most of China, reproducing biases that have plagued regional modeling efforts over East Asia for years. These errors matter far beyond academic bookkeeping. Regional climate models such as RegCM4 are workhorses for dynamical downscaling of global projections, assessing extreme heat and rainfall risks, and informing adaptation planning. If a model systematically runs too hot and too wet across the world’s most populous country, every downstream projection inherits that distortion.

Among the four configurations tested, one stood out. The experiment the authors call SEXPMF, which couples the Nogherotto–Tompkins microphysics scheme with the Xu–Randall cloud fraction scheme, delivered the best simulation of near-surface air temperature and precipitation across most of China. The improvements were quantitatively substantial. Compared with CTRL, SEXPMF reduced the root mean square error of two-meter air temperature simulations by more than 28.4 percent across all sub-regions of China, and cut the precipitation root mean square error by more than 14.5 percent in the Tibetan Plateau, Northwest China, and Northeast China. In Southeast China, where the scheme actually pushed precipitation magnitudes further above observations, the spatial correlation of the simulated precipitation pattern still improved by 49.0 percent, meaning the model got the geography of rainfall considerably more right even as it exaggerated its intensity.

The mechanistic explanation is where the study becomes genuinely fascinating, because it reveals how a seemingly local change in cloud physics can reorganize an entire regional circulation. The SEXPMF configuration generates more high-level ice clouds than the default setup. These icy veils intercept incoming sunlight, reducing the downward shortwave radiation that reaches the surface. Less solar heating at the ground means the widespread warm bias in surface temperature is alleviated. But the story does not stop at the surface. The altered heating profile creates what the authors describe as an upper-warming and lower-cooling vertical thermal structure, which strengthens atmospheric stability in the column above.

Over the Tibetan Plateau, that stabilization proved decisive. The plateau acts as an elevated heat source in summer, and spurious model heating there has long driven artificial convection and exaggerated rainfall. With the new cloud scheme damping surface heating, the modified thermal field triggered deep anomalous subsidence and moisture divergence over the plateau. Sinking air and diverging moisture are precisely the opposite of what convective storms need, so the model’s spurious convection was fundamentally suppressed and the overestimated precipitation reined in. In other words, a change in how the model represents ice clouds cascaded downward through radiation, surface energy balance, vertical stability, and finally the large-scale dynamics that govern whether storms can form at all.

Southeast China told the opposite story, and it is a cautionary tale about compensating errors. There, the modified thermal field produced anomalous upper-tropospheric divergence paired with low-level moisture convergence, a combination that drives strong vertical ascent and creates favorable conditions for cloud formation. The model responded by generating abundant clouds and, consequently, excessive precipitation. The authors attribute this over-simulation to compensating errors: improvements in one part of the system unmasked or even amplified biases elsewhere. The precipitation pattern became more realistic in its spatial structure, yet the magnitude drifted further from observations. It is a vivid illustration that model skill cannot be judged by a single metric in a single region, because the climate system’s internal compensations can trade one error for another.

Why should cloud parameterization choices matter so much in the first place? Clouds occupy a peculiar position in climate models: they form at scales far smaller than a model grid cell, yet they control the planet’s energy budget by reflecting sunlight and trapping infrared radiation. Microphysics schemes govern how water vapor condenses into cloud droplets and ice crystals, how those particles grow, collide, and precipitate, and how long clouds persist. Cloud fraction schemes, sometimes called macrophysics, determine how much of a grid box is covered by cloud, which directly controls the radiative fluxes computed by the radiation code. The two must work in concert, and the study demonstrates that their interaction, not either scheme alone, determines the fidelity of the simulation. A better microphysics scheme paired with a mismatched cloud fraction diagnostic can still produce a worse climate.

The findings carry practical weight for the regional modeling community. RegCM4 is one of the most widely used regional climate models in the CORDEX framework, which coordinates downscaling experiments across the globe, and it has been applied extensively over East Asia for monsoon studies and climate projection work. The identification of the Nogherotto–Tompkins plus Xu–Randall combination as the top performer over China gives modeling groups a concrete, evidence-based configuration choice rather than a default inherited from historical precedent. More broadly, the study underscores that the critical ingredient in simulating complex regional climates is the coupling among clouds, radiation, and dynamics, not the isolated performance of any single parameterization. A scheme that produces the right cloud amounts for the wrong reasons, or the right clouds without the right dynamical response, will ultimately mislead.

There is also a larger lesson about uncertainty in climate science. Cloud feedbacks remain among the largest sources of spread in climate sensitivity estimates, and the Intergovernmental Panel on Climate Change has repeatedly flagged them as a key uncertainty in projections of future warming. Studies like this one, which trace exactly how a parameterization change propagates through radiation, stability, and circulation to reshape regional climate, provide the process-level understanding needed to narrow that uncertainty. For China, where summer monsoon rainfall sustains agriculture for over a billion people and heat extremes are intensifying, the difference between a model that runs 28 percent too hot and one that does not is far more than a technical footnote. It is the difference between projections that can be trusted and projections that must be corrected. The Nanjing team’s work shows that sometimes the path to better climate information runs not through bigger computers or finer grids, but through the quiet, meticulous business of choosing the right equations for the clouds.

Subject of Research: Evaluation of cloud microphysics and cloud fraction parameterization scheme combinations in the RegCM4 regional climate model for simulating summer temperature and precipitation over China

Article Title: Impact of different cloud parameterization scheme combinations in the RegCM4 on the summer climate simulations over China

Article References: Li, F., Huang, A., & Gu, C. (2026). Impact of different cloud parameterization scheme combinations in the RegCM4 on the summer climate simulations over China. Climate Dynamics, 64(11), Article 448. https://doi.org/10.1007/s00382-026-08408-5

Image Credits: AI Generated

DOI: 10.1007/s00382-026-08408-5

Keywords: RegCM4, cloud parameterization, cloud microphysics, cloud fraction, regional climate modeling, China summer climate, Tibetan Plateau, precipitation bias, radiation-cloud coupling, Climate Dynamics, monsoon simulation, model evaluation