���Ϻ���APP

Do you check the weather forecast before heading out? Did you know that it is actually a numerical model, one that most people use every day? Few know that their daily forecast is not based only on equations. In fact, the development of weather predictions also involves leveraging observational data through a process called “data assimilation.”

Data assimilation is important in numerical weather forecasting because it enhances model accuracy by reducing initial condition errors that otherwise grow rapidly and affect the entire forecast. Data assimilation ensures forecasts remain relevant and accurate even as conditions change. Data assimilation also improves decision making through ensemble forecasting, which makes multiple weather predictions with slightly different starting conditions to estimate a range of possible outcomes and evaluate the likelihood of various weather events. This is significant not only in weather forecasting but also in long-term analysis for climate research. Without data assimilation, the estimated current state of the atmosphere would be unrealistic, leading to inaccurate and unreliable weather forecasts, impacting decision making in critical areas like disaster preparedness and mission readiness, and harming efforts to build community resilience. Data assimilation is a prerequisite for forecasting, and refining data assimilation methods is essential to producing more accurate weather forecasts.

The main goal of data assimilation is to provide accurate starting conditions for weather models. Each point in the model needs to know the initial state of the environment at its corresponding location, but observations are not always available for every location around the world. Most observations are taken at or near the Earth’s surface, while some observations come from weather balloons that provide vertical profiles, and many others are collected from satellites. While satellites cover large areas, they rely on assumptions to interpret the data. This results in a limited view of the atmosphere, especially above the surface and in remote areas. To fill in the gaps, data assimilation uses a previous model forecast as a starting point. Instead of just fitting observations to the model grid, data assimilation adjusts the forecast based on the observations. Common data assimilation methods include optimal interpolation, Kalman filters, particle filters, variational methods, and hybrid approaches. By combining observation data and model predictions using advanced statistical methods, data assimilation creates a more accurate picture of the atmosphere for the model to use.

Data assimilation systems are extremely helpful, but they can also be computationally intensive. For example, the National Weather Service (NWS) uses supercomputers that are more than 10,000 times faster than typical desktop computers. These supercomputers help produce daily forecasts in a few hours instead of several hours or days. Their impact is especially pronounced when used for hurricane prediction.

However, as the need for precise predictions grows with the more regular occurrence of extreme natural disasters like regional floods, the demand for high-resolution, frequently updated forecasts is increasing. In such circumstances, traditional data assimilation methods alone can become extremely computationally demanding and may not provide the timely or cost-effective results needed for preparation and response efforts.

Moreover, even with significant computing resources and extensive observational data, errors in models can still occur due to approximations in model physics or still-limited understandings of the relationships between current observations and future predictions. Machine learning (ML) and AI can significantly enhance data assimilation by detecting complex patterns and relationships in data that traditional statistical methods may not easily identify.

While some literature is showing AI/ML model improvements predicting extreme conditions at longer lead times (10+ days in the future) (Price et al 2025), traditional numerical models and data assimilation remain valuable and provide better physically interpretable explanations. AI approaches are highly sensitive to data coverage and quality, and despite ongoing expansions in observational networks, current observation density and quality are not yet sufficient to replace physics-based numeric weather models and data assimilation systems on a larger scale. These two approaches complement each other rather than conflict.

AI can enhance bias correction, better identify forecast uncertainty, improve data assimilation inputs, and integrate with physics-based constraints, ultimately creating hybrid systems that utilize both data-driven insights and fundamental atmospheric dynamics. This hybrid approach ensures forecasts that are more accurate, scientifically grounded, and practically useful.

As part of research and development at ���Ϻ���APP, we augment data assimilation and numerical weather forecasting with ML and AI techniques to enhance data quality and process efficiency, improve model bias corrections and prediction accuracy, and scale the model execution.

1.�������� Enhance data quality and data processing

Observational data often contain errors, gaps, and inconsistencies that degrade model performance. ML and AI can address these issues by automating data quality control tasks (such as labeling, cleaning, and error detection), resulting in more reliable inputs for numerical weather prediction (NWP). For example, ML-based methods can accurately classify different types of observations��(Jones, 2017), ensuring that spurious data are flagged or excluded.��In addition, AI can facilitate data assimilation by discerning the unique error characteristics of each observational��dataset, assigning��appropriate weights to improve initial conditions for NWP models.��This��can dramatically speed up and enhance analyses and forecasts at a reduced computational cost��(Keller & Potthast, 2024).��By improving both data integrity and the assimilation process,��the use of ML and AI provides a stronger foundation for downstream modeling tasks.

2.�������� Improve model prediction accuracy

Once the data are cleaned and better organized, ML and AI can��further��refine prediction accuracy by uncovering complex patterns and relationships that traditional methods often miss. By analyzing large volumes of historical and real-time data, ML and AI��can effectively��capture the initial state of the atmosphere, which is crucial for reliable NWP.��This approach��not only accelerates data assimilation but also��helps correct biases and fill coverage gaps, ensuring that models have access to a more complete and consistent dataset.��For instance, ML can dynamically adjust observational weighting based on data reliability, allowing high-quality observations to influence model initialization more strongly. In this way, ML and AI can leverage multiple datasets and known relationships to support better forecasts, leading to enhanced predictive��capabilities��and more timely weather insights for decision makers.

3.�������� Accelerate model execution

Many Earth system models, including weather forecasting models, are primarily implemented in Fortran, but most modern ML and AI models and libraries are developed in Python. While Fortran excels in scientific computing and complex numerical calculations, it lacks built-in support for automatic differentiation, creating challenges in integrating ML and AI methods and enabling hybrid models. Fortran also has more limited native Graphics Processing Unit (GPU) support, requiring additional tools or libraries��to fully utilize GPU acceleration.

Historically, rewriting these systems in other languages was considered burdensomely complex, leading to continued development in their original languages and making system modifications difficult. Now, with the help of generative AI, switching coding languages has become more feasible. For example, Zhou et al. (2024) utilized a large language model (GPP-4) to translate a photosynthesis model from the community Earth system model from Fortran to Python/JAX, resulting in a significantly faster runtime by utilizing GPU parallelization and parameter estimation via automatic differentiation. With generative AI's support, modernizing traditional weather models has become more achievable, offering faster performance and the ability to leverage recent advancements in computer science, thereby supporting novel cross-disciplinary collaborations.

Generative AI can enhance the data assimilation process by generating synthetic data to fill observation gaps. Unlike traditional machine learning, which relies on assumptions like linearity or Gaussianity, models such as generative adversarial networks and diffusion models produce realistic, high-resolution synthetic data that capture underlying nonlinear dynamics (Qu et al., 2024). These models use physical constraints to ensure data aligns with atmospheric dynamics. Incorporating synthetic data into assimilation frameworks helps achieve optimal initial conditions quickly, especially in regions with sparse data. This approach is valuable for time-sensitive operations like hurricane tracking, providing near-real-time data for faster assimilation.

���Ϻ���APP provides an AI-ready solution to enhance input data quality, strengthening data assimilation and forecast accuracy. We offer an open-source AI development toolkit named aiSSEMBLE™️, which supports efficient data storage ingestion, processing, and model inferencing for data assimilation. aiSSEMBLE™️ standardizes the design, development, and delivery of AI solutions throughout the engineering lifecycle, including data processing, model building, tuning, training, and secure operational deployment. This framework facilitates the integration and deployment of our AI-enabled data assimilation solution.

To enhance data assimilation, we are training a recurrent neural network (RNN) to approximate the background error covariance matrix using an approach similar to the National Meteorological Center (NMC) method (Chattopadhyay et al 2023). The NMC method is a classic technique in weather prediction that estimates forecast errors by comparing two different forecasts made for the same time, providing insights into the model’s uncertainty and creating better starting points for future predictions. In the NMC framework, historical forecasts valid at the same time but initialized from different lead times are compared to estimate a single, representative background error covariance matrix. By learning these relationships, the RNN can capture a more sophisticated picture of how uncertainties evolve in the model state.

Integrating this improved error covariance information into the weather model’s data assimilation process may provide more reliable initial conditions for each forecast cycle. These better-informed initial conditions lead to more accurate and reliable weather predictions and deliver faster and more cost-effective forecasts, ultimately aiding in better decision making to build a weather-ready nation.