Using parametric runs as training data for a machine learning algorithm

Dear Ladybug community,

I am currently in the process of setting up the framework for my master thesis. I need some input from you my fellow friends .

I want to find the correlation between the building envelope and floor plan in terms of daylight. For instance, using the VSC methodology and Daylight factor. What I can do is a parametric run of different room and shading types by including VSC analysis on the windows. I could then later train the model in a machine learning algorithm which can predict different outcome/solutions for buildings based only on VSC. After documenting the final solution, we can again do a VSC simulation and feed the properties to the ML algorithm for more accuracy down the line. This may also be included for Thermal comfort and radiation analysis later on. Using the VSC methodology for predicting daylight in a room might not be accurate because reflections and angle of rays are not included. Maybe this also needs some testing and implementation.

I am by no means expert on either climate analysis or machine learning. But curious and want to expand my knowledge in this field. Do you guys think this Is this feasible? Could it be done in some other way? I am open for input and suggestions :slight_smile:

Thank you, best regards Tobias.


This paper may shed some light on the topic for you:

‘Machine Learning Algorithm-Based Tool and Digital Framework for Substituting Daylight Simulations in EarlyStage Architectural Design Evaluation’ Radziszewski K., Waczyńska M.:

I am not an expert either but I think that VSC methodology you propose may be somewhat good for simple evaluations on bigger scale (urban planning) but as you mentioned, it wouldn’t take reflections and, what i belive is more problematic, zone (room) shapes and glazing placements into account. This does not apply if you would use typical Core+Perimeter layout or any other without internal partitions in daylit zones + same openings and window-to-wall ratios throughout the model.

I think you should also re-consider a metric of your choice. Daylight Factor is universal, which is both a ‘for’ and ‘against’ argument. It uses same CIE Overcast sky model, which is not dependant on location. This would mean that your data for ML is valid throughout the world but is not climate-based (as Daylight Autonomy, for example). If you decided to go for climate-based metrics you would have to iterate the process again for each location (and probably take other factors into account, such as window orientation for example).

An interesting topic, good luck with your thesis!

1 Like

Hi @Wujo.

Great work, thank you for interesting read.

How would you suggest to add additional window geometries? Would it be possible train the model based on 3d objects in JSON format?

I do agree with you on the choice of metrics and will consider climate-based. In addition it has more correlation with thermal comfort and overheating.

Best regards Tobias.