1. Introduction
ICMEs are among the most impactful solar phenomena for near Earth space. Expelled from the solar corona at speeds often between 20 and 2500 
$km/s$ Webb & Howard (Reference Webb and Howard2012), they can reach Earth in one to two days, driving severe geomagnetic storms, disrupting communications and navigation systems, and affecting space and ground based infrastructures Schrijver (Reference Schrijver2015). Accurate estimation of TT is therefore essential for early warning and risk mitigation.
Despite advances in physical and empirical modeling, TT prediction remains challenging. Traditional approaches such as manual catalogs or velocity distance relations are constrained by computational cost, and dependence on expert analysis, while the dynamic interplanetary medium introduces further uncertainties. Machine learning methods have been increasingly applied in space weather over the past few years Lugaz et al. (Reference Lugaz, Liu, Hapgood and Morley2021), but their progress remains constrained by the limited availability of datasets with essential variables and by the incompleteness of observational records.
These challenges underscore the need for flexible, data driven frameworks. SWAAT addresses this demand by integrating automated event validation with synthetic data generation guided by an evolutionary algorithm, which mitigates common issues in synthetic creation such as overfitting and oversampling, while ensuring low error. Its hybrid architecture combining evolutionary and generative models in the prediction stage with automated processing of GOES/XRS time series and SDO imagery in the validation stage enhances automation and strengthens predictive accuracy.
2. Background
A central challenge in space weather prediction is the scarcity of high quality, reliable data. Although solar missions have produced extensive observations, ICME events remain relatively rare, limiting the massive training of robust machine learning models. Existing catalogs, often based on expert manual validation, ensure quality but slow down the delivery of results.
Traditionally, event analysis has relied on IDL, while in this work we employ Python and multiple scientific libraries, with SunPy Barnes et al. (Reference Barnes and Christe2023) as the most widely used.
In astrophysics, the use of generative and evolutionary AI remains limited compared to supervised methods already applied in tasks such as galaxy classification, exoplanet detection, and solar structure segmentation. Generative models have been explored mainly for image enhancement or classification support, while evolutionary algorithms have been restricted to parameter tuning rather than dataset generation Zhang et al. (Reference Zhang, Wang and Zhang2024).
3. System Overview
The architecture of SWAAT (Fig. 1) is organized into two main processes: the Automatization Process and the Learning Process, each subdivided into functional stages.
• Automatization Process: Candidate events are analyzed by the ADAA module, which validates their significance using solar X-ray and imaging data. Non relevant events are discarded, while validated events trigger both the automatic report generation module and the extraction of the Vp.
• Learning Process: This process integrates two stages. The GESD expands the set of confirmed events by producing synthetic datasets optimized through evolutionary strategies. These enriched datasets feed into the Prediction Stage, where a polynomial regression algorithm estimates the ICME TT.
Architecture of SWAAT, showing the Automatization Process for event validation (top) and the Learning Process with synthetic data generation and prediction (bottom).
Figure 1 Long description
The flowchart shows the architecture of SWAAT, divided into two main processes: the Automatization Process at the top and the Learning Process at the bottom. The Automatization Process includes the Date of Candidate Event, ADAA for validation, and Automatic Report Generation. If the event is valid, it proceeds to Automatic Report Generation; if not, it is discarded. The Learning Process involves Original Confirmed Data, GESD for synthetic data generation, and a Polynomial Regression Algorithm for prediction, leading to the Final Prediction. Arrows indicate the flow and interaction between these components.
Together, these processes ensure a consistent workflow: ADAA provides reliable validation, GESD enriches the training base, and the prediction stage delivers the final TT forecast.
3.1. ADAA: Automation Component
ADAA is responsible for the early detection and validation of candidate solar events. It operates on two key data sources: GOES X-ray observations and SDO/AIA imagery.
GOES/XRS time series are analyzed in the high energy channel (0.1-0.8 nm), where rapid increases above a predefined threshold (M or X class) are flagged as potential candidates. Noise filtering, peak detection, and temporal derivatives are applied to discriminate genuine events from background fluctuations. In parallel, automated processing of SDO imagery defines a Region of Interest (ROI) covering 90% of the solar disk and identifies Points of Interest (POIs) as localized active regions. The temporal evolution of each POI is monitored, and a discrete derivative is computed to obtain a variability index.
An event is validated only when both criteria are satisfied: (i) the X-ray intensity exceeds the threshold, and (ii) POI variability exceeds the predefined threshold within the same time interval as the XRS peak. Valid events are then documented in automated HTML reports containing temporal plots and segmentation maps, while non events are discarded.
This dual condition process ensures consistent and reproducible validation, providing a reliable entry point to the SWAAT predictive pipeline.
Finally, when an event is validated, the Vp is derived following Salas-Matamoros & Klein (Reference Salas-Matamoros and Klein2015), by computing the soft X-ray fluence as the area under the GOES/XRS curve between the initial time (ti) and the maximum pear (tmax).
In Figure 2 the left panel illustrates the identification of solar flare peaks in GOES/XRS time series using a peak detection algorithm that determines the initial time (ti) and maximum peak at tmax. The right panel shows the extraction of Points of Interest (POIs) from SDO imagery through contour based analysis within the Region of Interest (90% of the solar disk). The temporal evolution of POIs is tracked, and a discrete derivative is applied to these measurements to compute a variability index, which serves as a criterion for event validation.
GOES/XRS time series with event initial time (ti) and maximum peak (tmax) (left), and SDO imagery showing the Region of Interest in green (90% of the solar disk) and detected Points of Interest in red (right).
3.2. GESD: Synthetic Data Generation through Generative-Evolutionary Methods
The GESD component is one of the most important elements of SWAAT’s learning stage. Here, a high quality synthetic dataset is generated to expand and diversify the original dataset, which are then used in the next step to predict ICME TT through polynomial regression. The synthetic generation process is guided by an evolutionary function that regulates the degree of the polynomial and ensures minimal overfitting, avoids synthetic overpopulation, and drives the model toward the lowest possible error. This capability is essential for overcoming the scarcity of data, a major limitation for machine learning models in predicting ICME TT.
The technical design and validation of GESD have been presented in detail in a previous study Meza-Obando et al. (Reference Meza-Obando and Calderón-Quesada2024). In this article, we provide only a functional summary and emphasize its integration into the operational workflow of SWAAT. For a full description, the reader is referred to the original publication.
4. Conclusions
The ADAA module has shown promising and consistent results in the automated detection and validation of solar events. It identified 
$\gt 95\%$ of significant X-ray peaks in GOES/XRS confirmed by manual inspection and correlated them with localized activity in SDO imagery. Within the defined ROI, the system successfully extracted POIs and quantified their temporal variability using a variability index derived from discrete derivative analysis, reducing validation time from 5-7 hours to 
$\lt 35$ minutes. This provides a robust and reproducible criterion for confirming candidate ICME events, closely matching the behavior expected from expert driven analyses.
A first detailed case study, based on a manually examined ICME event, confirmed that ADAA reproduces key observational features with high fidelity, both in X-ray dynamics and in the evolution of active coronal regions. However, due to the inherent complexity and time consuming nature of full manual analyses, the validation process is still ongoing. Additional events are currently being tested against carefully established manual baselines, ensuring that ADAA’s performance is benchmarked rigorously before being fully integrated into operational workflows.
The GESD module, responsible for TT prediction, expanded the training dataset from 86 real events to 260 synthetic+real instances, overcoming the scarcity of observational records. Models trained with this enriched dataset reduced the MAE in ICME TT estimation to 9.7 h, compared with 
$\gt 10$ h using other approaches. This demonstrates the scientific value of SWAAT: by uniting automated validation with informed synthetic data generation, it provides a reliable, reproducible, and operationally relevant framework for improving ICME forecasting.
5. Discussion
A key contribution of SWAAT is mitigating the scarcity of ICME datasets, defined only by TT and Vp, through synthetic data generation, avoiding the need for multiple or complex features. In this framework, generative models are guided by an evolutionary strategy, ensuring that synthetic datasets remain consistent while avoiding common issues in synthetic generation and reducing prediction error. This approach expands the dataset and decreases forecasting error in ICME TT estimation. Beyond space weather, the strategy also provides a flexible methodology for other astrophysical domains facing similar data limitations.
Despite its advances, SWAAT remains limited by its dependence on continuous X-ray data from GOES and solar imagery from SDO. Interruptions, resolution variability, or data gaps can compromise both detection and prediction. Although redundancy was considered at the design stage, the current system lacks multisatellite integration and robust strategies for handling missing data, highlighting clear directions for future improvement.
A promising development for SWAAT is its integration with the Radio Observatorio Santa Cruz (ROSAC), the first solar dedicated radio observatory in Central America. Since ground based radio flux measurements are strongly correlated with X-ray outputs, the same procedure for estimating Vp can be applied. ROSAC will therefore extend SWAAT’s observational base and enhance ICME prediction. Technically, this integration will require new pipelines and calibration routines, but it will also reduce dependence on satellites, strengthening the system’s capacity for near real time space weather monitoring.
