This article will cover various aspects of Weather Models, including:

Let's start with an example, Tomorrow at 2pm, the European weather model ECMWF may forecast the wind to be 15 knots from the East at the isle of Wight. One may wonder how such wind forecast is calculated in the first place and more generally how a numerical weather model works. Additionally another model like the American model GFS may forecast something different, making this confusing for anyone.

This article will dive into how weather models are designed and how they work. By understanding those concepts, the mariner will be able to better interpret numerical weather forecasts and compare the different models, with the goal of making better and safer decisions at sea.

Weather models often produce different forecasts for the same location, which can be confusing for sailors. For example, tomorrow at 2 p.m., the European weather model (ECMWF) might predict 15 knots of wind from the east at the Isle of Wight, while the American model (GFS) forecasts something different. But how are these forecasts generated, and why do models sometimes disagree?

This article explores how numerical weather models are designed and how they work. By understanding these concepts, sailors can better interpret forecasts, compare different models, and make safer, more informed decisions at sea.

Weather Models are a complex topic, so let's start with a simple comparison. Numerical Weather Forecasting is like "baking a cake".

Let's now break down in detail those four steps:

To forecast the weather in the future, you first need to know the current weather.
This is your starting point, and the weather model will calculate the change from this initial state. In Meteorology, this starting point is called the Initial Conditions and can be summarized as a snapshot of the atmosphere now.

Knowing the state of the atmosphere now means we need to measure all the weather parameters, such as wind speed, wind direction, pressure, temperature and humidity. This must be done everywhere, worldwide, not only at the surface but also at altitude. Plenty of instruments are used for this complex task, including land-based stations, buoys , ships, aircraft, satellites, and many more, as shown in the image below.

The image above is impressive, but the reality is the state of the atmosphere at a given time cannot be perfectly known, either due to missing data or the imprecision of weather instruments. The bottom line is that the inaccuracy of initial conditions is one of the reasons why numerical weather forecasting can not be perfectly accurate, and forecasts can not be correct 100% of the time.

So, to summarize, by not having perfect initial conditions, i.e. the perfect ingredients, you cannot make a perfect forecast, i.e. a perfect cake.

Making Initial Conditions is also a complex process, and different weather agencies create different initial conditions to represent the same atmosphere at the same time.
​
For example, the two weather models, GFS and ECMWF, even if they use many of the same instrument data to make initial conditions, come up with different initial conditions. This is one of the reasons why the GFS and ECMWF may differ. Also, as a side note, PredictWind uses those two different initial conditions to generate two different proprietary models:

Finally, the Initial Conditions are generated daily at specific times using UTC time. Typically, they are referred to as 0 UTC or 12 UTC.

The Initial Conditions are then used to forecast the weather a little bit later, referred to as the model time step, and this can be summarized in a single equation:

This equation starts at time equals 0, which is now, i.e. 0 UTC. The weather forecast, let's say 1 minute later, is the weather of the Initial Conditions plus a term F(A), which represents the combination of all the kinds of forcing that can occur multiplied by the time step (e.g. 1 minute) to equal the weather change during this one minute. Then, we keep going for the next minute and the next minute until we get a 10-day forecast.

The following paragraph will discuss function F, which is the physics equation that makes the heart of a weather model.

The atmosphere is too big and complex to simulate all at once, so scientists divide it into a 3D grid. Imagine slicing the atmosphere into tiny cubes, like a giant Rubik’s Cube. Each cube represents a small piece of the atmosphere, and the model calculates what’s happening inside each one.

The cube has a defined length, and it is the weather model resolution. It can be given in degrees or kilometres. For example, the ECMWF model has a resolution of 0.1° or 9 kilometres.

The weather model uses mathematical equations to predict how each cube will change over time. These equations are based on the laws of physics, such as Newton’s laws of motion and the laws of thermodynamics. Those equations are presented in the image below.

Those equations are intimidating, but they are essential as they are used by all weather models. It is unnecessary to understand those equations in detail, but a basic understanding of some terms may help the reader not to consider a weather model like a black box.

We have put below two videos and an article for the interested reader, but feel free to skip to section 1c.

The computer runs the simulation for each grid point and every time step, thus calculating how the atmosphere evolves over time everywhere on the globe. This process is repeated over and over, updating the forecast for the next time step until reaching the total forecast length (e.g. 10 days for ECMWF). The model output is usually simplified and distributed every hour (even if the timestep of the model is much shorter, around a few minutes). Very large computers are needed to run those models; see the image below.

Calculating a 10-day weather forecast takes a few hours, and this is one of the reasons why the forecast using initial conditions at 0 GMT is not available at 0 GMT. Generally, a few hours are required, and the GFS 0 GMT forecast is available at 4:40 GMT, and for ECMWF, which has a higher resolution, it is later at 7 GMT. This delay is also due to the time it takes to generate Initial conditions files, process the output, and distribute the forecast to PredictWind or other weather companies.

PredictWind has developed a tool to easily identify the current and next model update time so the customer knows how long he needs to wait to get the latest forecast.
​

For more information about model update times, see Model Update Times (GMT)

The forecast output is distributed in a format called a GRIB (Gridded Binary), which is organized in terms of:

See below a table summarizing this for some models:

PredictWind downloads those weather outputs from GFS and ECMWF twice daily (0 UTC and 12 GMT run). Models are also run at 6 UTC and 18 UTC, but the initial conditions used at those times have a lower resolution and are not distributed by PredictWind.

In section one, we presented global models that forecast the weather for the entire world. Those models are great for forecasting large weather patterns, such as a big low crossing the Atlantic Ocean and its associated wind.

However, due to their large scale, such models cannot capture localized weather patterns like a local sea breeze or the wind speed acceleration between two hills. Regional models have been developed to solve this. Those regional models zoom in on a specific area, typically a country, and forecast this smaller area at a higher resolution that allows them to capture local effects.

Regional / Mesoscale Models (Higher resolution, short-term forecasts)

PredictWind has developed two regional models called PWG and PWE 1 km. Click here to know more about those models.
​
These models are not restricted to a specific country but focus on the coastlines popular for sailing worldwide. Areas covered include all the USA, Europe, Australia, New Zealand coastlines, and much more.
​
For details, please see PredictWind Worldwide Coverage Map by Forecast Model.

The wind map below is for the location: Newport, USA. The map uses the split screen feature (toggle button highlighted in red) that allows you to put one model (GFS 25km) on the left and another model (PWG 1km) on the right.
​

PWG clearly shows how the Northerly wind interacts with the land and water channels, whereas GFS cannot give this information due to the low resolution.

Here is a great video to know more about model resolution:

So far, we have discussed deterministic weather models. These models take one set of initial conditions and give one forecast solution. This method is the most accurate and works well in the short to medium range, i.e. a few days. However, forecast accuracy decreases with time, so another type of weather model called 'Ensembles' can be used.

An Ensemble weather model takes a different approach. Instead of one forecast, it runs many slightly different simulations (called ensemble members), each with small changes in initial conditions and physics (e.g., ECMWF ENS). This helps reveal the range of possible weather outcomes, making predictions more reliable, especially for long-term forecasts and extreme weather.
​

Ensembles are used to track hurricanes that are very hard to forecast. Each yellow line below is a track of the hurricane centre for one ensemble member. Such a plot is sometimes referred to as a "spaghetti" plot. If the yellow tracks are grouped and tight, there is a strong chance the hurricane will follow this patch. If the yellow tracks are not grouped and diverge, there is a lot of uncertainty about the track of the hurricane.
​

PredictWind uses Ensemble for weather routing longer than 10 days, which extends beyond the ECMWF deterministic model.

If your weather route extends longer than 10 days, then the calculation of the weather route after 10 days will be done using one member of the Ensemble forecast. This is useful, for example, for an Atlantic crossing that will last longer than 10 days. The route will take you to your end waypoint in the Caribbean instead of ending in the middle of the ocean.

For short-term forecasts (up to 36 hours), high-resolution models like PredictWind’s PWG and PWE 1km, or local models such as AROME, HRRR, and NAM, along with global models like ECMWF and GFS, provide detailed wind and wave information. For longer-range planning (up to 7–10 days), global models including ECMWF, SPIRE, UKMO, and PredictWind’s own models help identify developing trends and assess forecast confidence.

We offer multiple models because there’s no single model that is consistently the most accurate. Performance can vary depending on the location and conditions—sometimes one model performs better, then another takes the lead. If all the models show similar patterns, you can have higher confidence in the forecast. If they differ widely, it’s a sign there’s more uncertainty in the situation.

It’s always best to review the highest resolution maps available—higher resolution usually means a more accurate forecast. Comparing the forecast to local observation stations and your own visual/instrument data can also help. For example, if PWG is predicting 25 knots from the south with rain, and that’s exactly what you’re experiencing, then it’s likely doing a good job and should be trusted more in the short term.

You can check the model accuracy ratings from a national weather centre here:
👉 Model Validation Technical Report

With only 1–2 models, it can be hard to know which to trust, but with access to up to 9 models, it’s easier to spot consensus and make more informed decisions.

To better understand what each model represents, check out this help article:
👉 PredictWind Model Terminology

While on passage, we recommend updating your weather route at least twice a day. You can also monitor live observations in your area to see which model is currently tracking best with actual conditions.

You should be checking the forecast at least twice a day. PredictWind’s global models update every 12 hours, and some of the regional models refresh more frequently. For a full list of update times, check out this article:
👉 Forecast Update Times (GMT)

Ensemble models, such as ECMWF ENS, help assess uncertainty and identify possible weather trends for longer-term planning beyond 7-10 days. If sea conditions are a primary concern, wave and ocean models like WW3 (WaveWatch III) or RTOFS offer detailed forecasts on wave heights, ocean currents, and sea surface temperatures.
​

Several key weather parameters should be monitored when planning a marine journey. Wind speed and direction are crucial, affecting sailing efficiency, fuel consumption, and vessel stability. Wave height and period determine how rough the sea will be, while ocean currents and tides influence navigation and fuel efficiency. Storms and severe weather, including cyclones and squalls, can pose serious risks, making it essential to track their development. Visibility and fog are also important factors, especially when navigating through congested or narrow waterways.
​​

When applying forecast data, it is best to compare multiple models to identify trends and inconsistencies. Ensemble forecasts should be used to assess uncertainty, particularly for long-distance journeys. Since marine conditions can change rapidly, frequent updates are necessary to ensure safe navigation. Adjusting routes and timing based on forecasted wind, waves, and storms helps optimize both safety and efficiency.
​

By using the right weather models and monitoring key parameters, sailors can make informed decisions, minimize risks, and ensure smoother, more efficient voyages.

As we have seen since the beginning of this article, Numerical Weather Prediction uses a fixed set of rules, i.e. the physical equations of meteorology, to calculate how the weather changes.
​

Nowadays, Artificial Intelligence (AI) is taking the world by storm, and it affects many domains, including weather forecasting. Artificial intelligence Weather Prediction uses an entirely different approach than numerical weather forecasting.
​
In an AI model, there are no physics equations or any rules of meteorology. Instead, the AI model is first fed with a lot of historical weather data, which enables the model to learn how to forecast the weather. This may be a bit confusing, so let's take a simple analogy to grasp the basic concept of the AI weather model.

Let's take the example of an old fisherman who leaves the dock every day at 5 am to go out fishing. With all his knowledge of going out to sea for the past 40 years, if you meet him on the dock one morning and ask him if it is going to be windy today, he may say, "10 knots from the North-East by 10 am when I am back at the dock, I reckon". Without running any numerical weather model in his head, he instead intuitively compares the weather now to all similar days he experienced in the past. Simply put, his experience enables him to make a forecast based purely on historical data he lived through.

AI weather forecasting is a bit similar but with some noticeable differences in scale. Historical weather data is made up of trillions of data points, and the AI model is made up of millions of parameters. Additionally, the AI model uses spatial patterns in observed data to project future weather conditions. Let's explain this in more detail now.

ERA5 is a global reanalysis dataset produced by the European Centre for Medium-Range Weather Forecasts (ECMWF). It provides hourly weather and climate data from 1950 to the present, covering the entire planet with a spatial resolution of about 31 km. ERA5 combines observations from satellites, weather stations, and buoys with advanced numerical models to reconstruct past weather conditions accurately. It includes key variables like temperature, wind speed, air pressure, precipitation, and sea surface temperature.

What is important to realize is how large the ERA5 data is. We are talking about 20 million gigabytes. Now, let's see how this large dataset can be used to train the AI weather model so it can forecast the future.

Weather data is typically represented in a weather map, which is nothing less than an image. Because of this, the type of AI used for weather prediction is the same as image processing.

A Vision Transformer (ViT) works by splitting the weather map into small patches, like cutting a photo into puzzle pieces. Each patch is converted into numbers and treated like a word in a sentence. A transformer model then looks at all patches together, using self-attention to find relationships between distant areas—like spotting a storm forming across regions. This helps AI models understand patterns in weather data more accurately and quickly. ViTs analyze the whole picture simultaneously, making them great for forecasting complex weather systems.

Below is how a Vision Transformer will take an image, break it down into 9 tokens and use AI methods called attention to find some patterns in an image. Similarly, a weather map could be used to find patterns like a cold front, a cyclonic convergence for deep low systems, etc.

The key concept is we need to train the model using historical weather data. Only when the model is trained can we use the model to forecast new weather data to predict tomorrow's weather.

To explain the training mechanism, let's take one single weather map, which is the air temperature on January 1st 2000, at 4 am provided by the ERA5 data set, see the image below - it does not represent the real data but is provided as a visual aid.

Using only image4am, the AI model will do a good forecast if it can generate the air temperature 6 hours later at 10 am. Since the ERA5 database also has the 10 am map in its database, the AI model will be able to compare its forecast with the image in the database and potentially learn from its errors to improve and learn.

At the beginning of the model training, the model will be given image4am and it will create his first imageForecast10am, and the forecast will in fact be really bad. The forecast image may be a uniform map of 20°C all over the world or maybe a map like the one below, which is slightly better but not even close to the real map image10am stored in ERA5 database.

Even if the first forecast done for the AI model from 4 am to 10 am is terrible, the model can compare ImageForecast10am to image10am stored in ERA5. The model will know where the forecast was wrong and by how much eg. a temperature error of +2° over London. Knowing all those errors everywhere on the map, the model will adjust its parameters to minimise the temperature errors.
​
Then the same image Image4am will be given into the model another time, and the forecast will be slightly better but still pretty bad overall. New errors will be calculated again, and the millions of parameters will be slightly adjusted again. By repeating this numerous times, the model will refine its parameters and get better and better. We will end up with a trained model, ready to be used to forecast air temperature using new data, which is an air temperature map today so an accurate forecast can be generated.

Here are some additional relevant points regarding AI weather models:

To conclude, an AI weather model can be imagined as a control panel with millions of knobs that can be adjusted during the training, so it does a good job forecasting the historical datasets. Such a panel can be visualized in the picture below.

Now that we have presented the core concepts of AI weather model, let's have a tour of what models are available at the moment in early 2025.

Here’s a comparative table of the most popular AI-based weather forecasting models, including Fengwu, GraphCast, FourCastNet and Pangu-Weather.

A question you may wonder is: Are AI Weather Models better than Numerical Weather Models?

This is a tricky question to answer, so we put below the answer provided by Chat-GPT, which is probably skewed toward AI weather models. So, let the reader be the judge!

AI weather models are improving rapidly, but traditional Numerical Weather Prediction (NWP) models still have key advantages in some areas:

To conclude, PredictWind is always looking for innovation and new technology to provide our users with the world's best weather technology. In 2025, PredictWind will add AI weather models to its offering of existing Global and Regional Numerical Weather Models. So stay tuned for more!