The AI-powered app that helps predict UK winter climbing conditions.
To continue and use this service you must agree to the disclaimer.
By deleting your account you will cancel your annual subscription renewal and lose out on all the benefits:
✔ 6500+ daily forecasts
✔ Conditions summary
✔ Favourites
✔ Recommendations
✔ Improve predictions
✔ Recent reports
✔ Only £11.99 per year
Are you sure you want to delete your account and unsubscribe?
These predictions are for the ethical condition of winter climbing routes. Importantly, they are not predicting the risk of approaching, climbing and descending from these routes. Please see our help section for further details.
Sign Up
Terrible
Poor
Marginal Reasonable Good
% The probability that the prediction is correct.
↑ The remaining probability is leaning toward better condition.
↓ The remaining probability is leaning toward worse condition.
Together these show both if and how the AI is unsure on a prediction.
To help the AI learn and improve future forecasts you can tap any prediction underlined with blue dots e.g. and feedback observed conditions.
The app predicts the ethical condition of winter climbing routes throughout the UK. Importantly it does not predict the risk of approaching, climbing and descending from these routes.
There is a long tradition of ethics for winter climbing in the UK with the community establishing guidelines such as:
Winter Climbing Forecasts observes the weather patterns over the last several weeks and then uses artificial intelligence to forecast which routes are likely to meet the above. This is tailored to the particular style, location and speed that each individual route comes into condition.
There are over a thousand routes and growing on the app. Each of a particular style:
For further information please see these guides on winter climbing ethics:
No. The variations on how climbers could approach and descend from a particular crag are endless, which makes it impossible to factor these into the predictions. For example, should the app consider going over the top or all the way round to Beinn Eighe Triple Buttress? Both have different aspects and avalanche potential.
On certain style of routes, such as snow gullies, the app does look for the snow to settle for the best ethical conditions. However, this is not a statement on whether it is safe to climb with regards to avalanche. As an example, the app lacks the specialism to know a persistant weak layer is within the snowpack nor be able to tailor to what a climbers risk tolerance to this is.
It is essential to use resources such as Scottish Avalanche Information Service to make this judgement. We highly recommend using their expertise and in-the-field knowledge as part of your planning. We provide links within the webapp to help you with this.
No. Each individual climber will have their own acceptable risk level, which makes it impossible for the app to make a judgement for every user. So it is essential to use resources such as Mountain Weather Information Service to you make this personal decision during the planning stage. We provide links within the webapp to help you with this.
There are some fail-safes within the code for very extreme weather. For example, any style of route will be predicted terrible if wind speeds go over 100mph. This would not be a statement on whether it is safe below this threshold, as the vast majority of the community will likely have an acceptable risk level significantly below such an extremely high wind speed.
There are also subjective hazards relating to weather such as poor visibility and a climbers ability at navigation, which again would make it impossible to tailor the predictions to every user.
No. The difficulty of a climb and a climbers ability to climb it is a subjective hazard that would be impossible for the app to assess for each user. In addition, objective hazards such as loose blocks, poor belays etc would be down to an individuals acceptable risk level.
Ultimately there would be a conflict between predicting the safest conditions and recommending routes plastered in rime, snow and ice to be in acceptable ethical condition.
Winter Climbing Forecasts is just one piece of the jigsaw that is the planning process before a trip. It is to be used with other useful planning tools such as:
We provide links within the app to some of the above to help with your planning. Note that we hold no responsibility for the content of these external sites, and providing a link does not constitute an endorsement.
In addition, we recommend reading this series from Plas Y Brenin and UKClimbing that looks at the planning process in more detail:
The final article in this series covers situational decision making, which is a good read on why forecasting tools such as this app can never be a substitute for your own dynamic risk assessment out on the mountain. With the app being based on weather forecasts, these only need to be slightly inaccurate to have an impact on Winter Climbing Forecasts own predictions.
Anyone who uses the service must first agree to the disclaimer, which can be read again here.
Non-technical Overview
The field of Artificial Intelligence (AI) can be traced back to the 1950s where the term was coined at a research conference. Initially
the techniques used involved a rules based approach to reason and solve problems. These Symbolic AI
methods dominated through the 1960s and 1970s and are still used in many expert systems today.
As an example, turfy mixed routes consider the following rules:
The 1980s saw the emergence of Machine Learning. This takes a different approach by feeding the algorithms labeled data, which the AI then learns to spot patterns with and then make predictions off new data in the future. With huge improvements in the amount of available data, plus more powerful hardware and refined techniques, the 2010s have seen an explosion in the potential of AI. These breakthroughs are especially down to a sub-field of Machine Learning known as Deep Learning. Here the AI effectively mimics the structures of a human brain to make deeper insights into the training data it is provided with.
Winter Climbing Forecasts combines both the above techniques to continually improve its predictions. Please see our S01E07 Rise of the Machines interview on the Vertical Voice podcast, if you would like to find out more.
Technical Overview
Datasets
The model needs to learn the relationship from weather and route features to the five condition labels.
We pre-process the weather dataset with a sliding window of seventy-days and then label from two sources - predictions from our rules based symbolic AI and feedback from users via the observed conditions feature.
At the start of the sixth season these datasets stand at 19.8 million and 37.9 thousand instances respectively.
Model
We use a Transformer decoder with RMSNorm, RoPE, SwiGLU and Stochastic Depth, then scale the model in accordance to the Chinchilla laws.
We find that using self-attention on weather features and cross-attention with the route features works well in this use case.
As does reducing the key_dim to increase the number of attention heads.
Training
Initially the model is pre-trained on the entire dataset of symbolic AI predictions during the off-season.
This is then followed by fine-tuning the model on the latest user feedback dataset three times per day.
Both stages use the Lion optimiser with warm-up and cosine decay, but with different hyperparameters.
Evaluation
The core metric we use to evaluate the performance of our model is 100 * (1 - Normalised MAE), which gives a more nuanced scoring of the ordinal classes.
At the start of the sixth season this gives a score of 91.25% ± 0.03σ on the 2.7 thousand instance test set.
Inference
Three times per day the latest weather forecasts for each location are retrieved and fed into the latest fine-tuned model to make the 26.6 thousand predictions needed for the next fifteen and previous five days.
To assess the confidence in these predictions we use Monte Carlo sampling with 100 forward passes.
Infrastructure
The app is hosted on a Debian server with a Nvidia RTX 3060 GPU and the backend is written in Python.
We use Dask for efficient parallelised preprocessing of the datasets and Tensorflow for the deep learning aspects.
Our server is powered by the green energy provider Ecotricity.
If you have any further questions, then please get in touch: adam@winterclimbingforecasts.co.uk
Winter Climbing Forecasts was created by Adam Godwin.
He is a keen winter climber who moved to Scotland from the Lakes over a decade ago. Highlights here include a winter traverse of the Cuillin Ridge and many of the classic Nevis ice climbs such as Point Five, Orion Face Direct and Minus One Gully. Outside of the UK he has done winter climbing trips to the Polish Tatras, European Alps, New Zealand Southern Alps and Peruvian Andes. He also dabbles in a bit of running and has completed the Bob Graham Round.
He can be contacted on: adam@winterclimbingforecasts.co.uk
OpenTopoMap contributors and © OpenStreetMap contributorsYes. Rather than a native app available on the App or Play Stores the tool is installable as a Progressive Web Application. This can be installed from the default browser on Android and iOS mobile device as follows:
on Android Chrome or 
on iOS Safari
The coloured pins across the map give a useful summary of what the best prediction for any route is on that particular crag in the coming days. They are a useful way of narrowing down suitable routes to climb.
For example, the pin 
would mean that one of the routes on that crag could be in Good condition within the next five days.
Tapping the pin, followed by a tap on the name of crag reveals the more detailed five day forecast for each individual route.
In prime time conditions it can be difficult to narrow down suitable climbs by browsing the map alone. To help with this there is a Filter option.
This displays any crags that have routes predicted as or on the dates, the grades and the style selected.
Select your options then tap Apply.
Once done with the filter you can choose Reset to go back to the original map view.
The search 🔍 box at the top of the page can be used to find a specific mountain, crag or route that you are interested in climbing.
Typing within this reveals a drop down of matches, which is refined as each character is pressed. Keep typing the name until its displays, then tap to go to that location.
As climbers we all have dream routes that we are waiting patiently to come into condition. The Favourite option allows supporters to quickly display the forecast for each of their dream routes.
To add a route to your favourites list simply tap its name in the detailed day-by-day view followed by Add Favourite.
A route can be removed by tapping its name in the favourite view and then tapping Remove Favourite.
To help you find climbs there is a personalised recommended routes feature included for supporters. This looks at your favourites list and recommends similar climbs with a preference for those that are currently coming into condition.
From the Favourites menu you can be presented with a "you may also be interested in" option, but seeing this depends on what routes you already have in your favourites. You can browse these recommendations by tapping the left and right arrows, then to add a recommended route simply tap on it and then on Add Favourite.
As a supporter you can tap on the current days prediction (or the two days before) and let it know what you saw out on the mountain.
Under the crag or favourites view tap any prediction underlined with blue dots e.g. to reveal a popup where you can feedback an observed condition.
This goes directly into the AI making it learn and improve the accuracy of future forecasts. This is the feature that we are most excited about, as a whole community using this over the long-term will make Winter Climbing Forecasts a very powerful app indeed!
At Winter Climbing Forecasts we are continually looking to improve the app and its predictions. This can be through the observed conditions feature or now through our new written feedback tool. Ideal if you have more detailed feedback to give.
If you would like to suggest improvements to the tool, then as a supporter:
This feature helps you narrow down which general style of winter climbing may be coming into condition. Styles include Mountaineering Ridges, Snow Gullies, Snowed Up Rock, Turfy Mixed, Icy Mixed, Water Ice, Snow Ice Gullies and Snow Ice Faces.
It provides a concise graphical overview of the amount of routes have been predicted to be in condition over the previous five days - and importantly the next fifteen days. With the line chart showing the percentage of each particular style that have been forecast either Reasonable or Good. Styles can be toggled on and off the chart by tapping them on the legend.
By default the conditions summary is for all of the UK, but using the buttons below the chart can be filtered down further to the following areas:
The model's summary is based on weather forecasts, so its reliability is directly dependent on the underlying data. As is typical with weather forecasting, accuracy decreases the further into the future it predicts, meaning a fifteen-day outlook will be less reliable than a shorter-term one. Despite this, the summary still gives a useful general sense of upcoming conditions.
We hope the app is of use to the community. If you do find it useful, then please sign up.
To signup:
This will take you to an external page hosted by Stripe where you can enter you payment details and setup a subscription. Upon a successful payment you will be sent back to Winter Climbing Forecasts and automatically signed into your account.
To signin to your account:
The website will remember you are signed in automatically. Unless your browser is configured to delete cookies when closed, which will require you to signin each time you access the website.
To signout of your account:
To reset to your account password:
A reset code will be sent to your email address. Copy this into the appropriate box then enter your new password and tap Reset.
To delete your account and unsubscribe: