Accurate atmospheric data is essential for weather forecasting, aviation safety, and climate research.

However, real-world data collection methods such as ra-diosonde launches are limited by high operational

costs, sparse temporal availability, and restricted geographical coverage. To address these challenges, this

paper proposes AtmosGen (Atmospheric Synthetic Data and Image Generator), a condition-aware syn-

thetic atmospheric data and image generation framework that combines numerical data synthesis with

atmospheric image generation. The system utilizes historical radiosonde data to generate realistic synthetic

atmospheric parameters, including temperature, pres-sure, humidity, wind speed, and altitude, using

machine learning-based generative mod-els. In addition, a conditional image generation model is employed

evaluation. The proposed approach reduces dependency on continuous real-time data acquisition while

providing scalable, diverse, and scientifically consistent datasets. This framework is particularly useful for

aviation simulations, machine learning model training, and atmospheric research where large labeled

datasets are required.

Weather Simulation

Accurate atmospheric data is essential for aviation safety, weather forecasting, and cli-mate research.

Aircraft operations rely on vertical atmospheric profiles such as temper-ature, pressure, humidity, and wind

speed, typically obtained from radiosonde observa-tions. Although radiosondes provide reliable upper-air

atmospheric datasets can reduce dependency on costly balloon launches while supporting scalable research

and aviation analytics. However, generating scientifically consistent synthetic atmospheric data that

preserves physical realism and statistical similarity to real observations remains a significant challenge.

This paper proposes AtmosGen, a Synthetic Atmospheric and Aviation Dataset Gen-erator using historical

radiosonde data as the reference source. The system applies machine learning-based generative models to

produce realistic atmospheric profiles, including temperature, pressure, humidity, and wind speed across

altitude levels. A compatibility and comparison model evaluates statistical similarity between original and

generated datasets using correlation metrics and error analysis. The framework also supports condition-

aware mapping between generated atmospheric parameters and corresponding weather representations.

Supporting Parameters – Aviation & Environmental Factors

To enhance realism, the system integrates contextual aviation-related parameters, in-cluding:

Visualization & Decision Support

Outputs include synthetic dataset previews, statistical comparison results, correlation graphs, and

atmospheric trend visualizations in CSV/JSON formats to support down-stream research and simulation

tasks.

Traditional atmospheric data collection systems primarily rely on radiosonde-based ob-servations for

obtaining upper-air measurements such as temperature, pressure, humid-ity, and wind speed at various

altitude levels. Radiosondes, which are balloon-borne instruments, have long served as a reliable source of

vertical atmospheric profiling for aviation safety, weather forecasting, and climate research. These

Furthermore, many remote regions and oceanic areas lack radiosonde stations, resulting in geographical

data gaps. While satellite systems provide broader coverage, they often lack the high-resolution vertical

profiling accuracy offered by radiosondes. Numerical reanalysis models depend heavily on available

observational inputs, and in-accuracies in input data can propagate through the modeling system. Another

limitation is the dependence on real-time physical data collection, which restricts scalability for applications

that require large, continuous, and diverse datasets. The absence of syn-thetic augmentation mechanisms

limits the ability to simulate rare atmospheric events such as turbulence spikes, extreme wind shear, or

sudden pressure changes. Moreover, traditional systems do not provide built-in statistical compatibility

validation when in-tegrating generated or interpolated data, which may lead to inconsistencies in advanced

modeling environments.

and oceanic regions. Existing alternatives such as satellite data and numerical reanalysis models provide

broader coverage but often lack the vertical resolution accuracy offered by radiosondes. Additionally,

current systems lack an integrated mechanism to generate statistically validated synthetic datasets ca-pable

of preserving physical realism while supporting scalable data augmentation. The primary objective of this

research is to develop AtmosGen, which leverages historical radiosonde observations to produce realistic

and scalable synthetic atmospheric profiles while preserving statistical distributions and physical

interdependencies. The system in-corporates aviation-related contextual variables such as turbulence

effects, wind shear patterns, and seasonal variability to enhance simulation realism. A key objective is the

development of a compatibility and comparison model using MAE, RMSE, and cor-relation analysis to

validate generated data against real observations. Ultimately, the proposed system aims to reduce

parameters, improving the accuracy and reliability of atmospheric condition analysis. Their study

demonstrated that combining multiple atmospheric indicators enhances the precision of environmental

modeling and supports more reliable atmospheric data interpretation, highlighting the importance of multi-

variable atmospheric modeling in synthetic dataset generation systems.

Gidey and Mhangara (2025) analyzed long-term atmospheric and environmental temperature relationships

using satellite-based datasets and demonstrated strong correla-tions between atmospheric conditions and

surface temperature variations. Their study emphasized the importance of accurate atmospheric data

modeling for predictive anal-ysis and environmental monitoring, supporting the need for scalable data

generation frameworks that preserve statistical relationships between atmospheric parameters.

Garai et al. (2022) investigated the relationship between rainfall, temperature, and envi-ronmental

Londhe et al. (2023) studied atmospheric and environmental variability across differ-ent climatic zones and

observed strong seasonal and regional variations in atmospheric parameters, demonstrating that modeling

systems must account for temporal variability to improve prediction accuracy and simulation reliability.

Awasthi et al. (2023) examined climate sensitivity and atmospheric variability and found significant

relationships between atmospheric conditions and environmental changes, highlighting the importance of

stability, underscoring the need for accurate atmo-spheric data modeling for environmental monitoring.

Ravuri et al. (2021) proposed a deep generative model for skillful precipitation nowcast-ing using radar

data, demonstrating that generative adversarial networks can produce physically consistent meteorological

outputs that outperform traditional deterministic forecasting approaches. Their work directly supports the

use of GAN-based architec-tures for realistic atmospheric data synthesis.

Stengel et al. (2020) investigated the use of GANs for super-resolution of climate model outputs, showing

that generative models can preserve physical constraints and statistical distributions of atmospheric

variables even when trained on sparse observational data. This work provides methodological grounding

for physics-aware generative modeling in atmospheric contexts.

Price et al. (2023) presented a hybrid neural network and general circulation model framework for

GANs, VAEs, and regression-based models are trained on historical data to learn temporal and spatial

atmospheric patterns, enabling simulation of diverse avia-tion and environmental scenarios including rare

or extreme atmospheric conditions.

+40◦C, pressure from 990 hPa to 1040 hPa, and humidity from 30% to 95%. Each record contains seven

features: al-titude (m), temperature (◦C), pressure (hPa), relative humidity (%), wind speed (m/s), wind

direction (degrees), and dewpoint (◦C). The dataset was generated using the Inter-national Standard

Atmosphere (ISA) physics model with realistic stochastic variation, incorporating a lapse rate of −6.5◦C per

1000 m, the barometric pressure formula, ex-ponential humidity decay with altitude, and jet stream wind

patterns and produce synthetic datasets. Descriptive analysis studies statistical distributions, correlations,

and altitude-based variations. Two generative architectures were implemented, trained on the full 50,000-

record dataset using a Google Colab T4 GPU and TensorFlow 2.x, and evaluated against the physics-based

ISA baseline.

1024 neurons) with LeakyReLU activations (α

a 100-dimensional Gaussian noise vector as

uses layers of 512, 256, and 128 neurons with LeakyReLU activations and Dropout (rate 0.3), concluding

assessment, and dataset consistency verification.

The GAN achieves the strongest overall performance, with Pearson correlations exceed-ing 0.99 for

temperature and pressure and above 0.96 for humidity and wind speed. The VAE performs competitively

evaluated on a shared test set predicting tem-perature from altitude.

Model B achieves R

humidity clipping to [0, 100]%, wind speed clipping to [0, 200] m/s, and the thermodynamic constraint that

dewpoint must remain below temperature. Despite these constraints, occasional inconsistencies were

observed in mid-tropospheric temperature profiles and in wind speed generation above 20 km, where both

models produce smoother profiles than physically expected, likely due to limited training diversity at high

altitudes. For extreme surface conditions at the boundaries of the training distribution, generated outputs