DesignSense: from a design space to a decision

The first three steps get the data into DesignSense. The other four are the tasks themselves, and you will move back and forth between them: what you learn in one often sends you back to an earlier one.

DesignSense does not upload your data. It is read in your browser, and your work (sets, shortlist, ratings, notes) is saved in that browser too.

A study is a table with one row for each design option. Its columns are the inputs that produced the option, such as a width, a glazing ratio or an orientation, and the outputs that were measured or simulated for it, such as cost, daylight or cooling load. Each option can also have a picture and a 3D model, stored as files next to the table.

DesignSense reads the format that Design Explorer introduced, so any tool that can make a study for Design Explorer can make one for DesignSense.

3D models

3D models are optional. DesignSense can show .glb, .gltf and .obj files, and the .json scenes that Design Explorer uses.

• A Design Explorer study with a threeD column of .json scenes loads as it is.
• Rhino 8: select the geometry, choose File ▸ Export Selected, and pick glTF from the file-type list.
• Rhino 7: install McNeel's glTF exporter from the Package Manager. Alternatively, export Wavefront (.obj) and tick Map Rhino Z to OBJ Y so that the model stands upright.
• Grasshopper: bake the geometry into Rhino, then export it as above. Export one file per option and give it the same name as that option's image.

A .glb is the safest choice because it is a single self-contained file. Add a model3d column to data.csv that contains each option's file name.

A study is one folder:

my-study/
├── data.csv          one row per option
├── 0001.png          the picture the row names
├── 0001_2.png        a second view of the same option (optional)
├── 0001.glb          its 3D model (optional)
├── 0002.png
└── …

data.csv starts with a header row, followed by one row per option. This header comes from one of Design Explorer's own studies, which loads without changes:

in:Depth,in:Height,in:Orientation,in:WWR,in:SHD,out:Cooling[kWh],out:Heating[kWh],out:Lighting[kWh],out:DA,img,threeD

What the column names mean

DesignSense works out what a column is from its name. The prefix is not shown in the interface.

Names are not case-sensitive. Colibri writes in: and out:. min: and max: are specific to DesignSense, and you can set the same goals after loading without editing the file. DesignSense writes custom: and weighted: columns itself when you download a table, so that your own criteria are restored when you load that table again.

What the cells hold

• Numbers become axes that you can filter, sort and colour by. A column counts as numeric when every filled cell in it is a number.
• Text is shown in the inspector and in Compare, and you can sort and colour by it. It does not become an axis.
• Empty cells are treated as missing values, not as zero.
• Picture and model cells contain either a file name (0001.png), which DesignSense looks for in the same folder as data.csv, or a full link (https://…/0001.png). Pictures can be .png, .jpg, .gif, .webp, .bmp or .svg.
• Commas, tabs and semicolons all work as separators.

Warning

A column called image_url or picture is not read as a picture. Rename it to img.

Coming from Design Explorer

• in:, out: and img mean the same thing, so a Colibri folder loads as it is.
• The threeD column and its .json scenes are read.
• A Description column, which Design Explorer's samples use, provides the option names.
• A settings file in the folder is ignored. You set goals, colours and layout in DesignSense.
• Google Drive links do not work. See the note under Generate.

Load Data is in the toolbar and on the welcome screen. It gives you three ways to open a study. Open Load Data

The Load data dialog, with its three sources: a prepared dataset, a link, and a folder or CSV file from your computer.

Checking how the columns were read

The first time you load a study of your own, the Your columns dialog opens and shows how DesignSense read each column. There you can change a column to an input, an output or neither, set which direction is better for an output, and choose the column that names the options. DesignSense remembers these choices for that study.

Your columns for the Windows study. It has five inputs and eight outputs, four of them with a goal, plus the columns that hold the pictures and the 3D models.

When a load goes wrong

All the views show the same options and stay in step: when you filter, select or inspect in one view, the others update. Help ▸ Take the tour shows each view in turn in the app.

Colour. You choose one column to colour by, and every view uses it: the lines, the dots and the border of each gallery card. Click an axis title to colour by that dimension, and pick a palette from the menu next to the colour key. Four other colours are fixed and always have the same meaning:

Lite and Expert. DesignSense starts in Lite mode, where you narrow the options with filters. Switch to Expert with the toggle in the toolbar. In Expert you build a selection step by step instead of filtering, and you also get clusters, the Analysis panel and layout editing (watch Expert mode). Everything below works in Lite unless stated otherwise.

The next four sections cover the four tasks. You will rarely do them once and in order. What you learn while deciding may send you back to narrow in a different way, and sharing a result often starts another round. The examples all use the 648 window designs shown at the top of this guide. Imagine an architect who has to choose a few of them to show a client.

Find the patterns, relationships and performance trade-offs in the design space. Compare alternatives and weigh their strengths and weaknesses to learn which criteria matter.

All 648 options, coloured by cooling. One option is inspected, so its picture is open on the left and its line is drawn thicker across the axes, with its values. In the scatterplot, the dashed line connects the best trade-offs between daylight (DA) and cooling. The overview next to it shows the other pairs.

Try it: Open the Windows study · Tour this task · Watch this part

• Patterns. The lines in the parallel coordinates show where options are concentrated and where there are few. Drag an axis title to put two axes next to each other. If the lines between them run roughly level, the two dimensions rise together. If the lines cross, one goes up as the other goes down.
• Relationships. The scatterplot shows one pair of dimensions in detail. The Overview shows every pair at once, which helps you spot the pairs worth a closer look. Colour by a third dimension to see what else changes with them.
• Trade-offs. After you give each output a goal, arrows on every axis show which direction is better, and the scatterplot can draw a line through the best trade-offs for the pair it shows.
• Form and performance together. Sort the gallery by a column to see how the designs change as that value changes. When you inspect an option, its line and its dot are highlighted in the charts, so you can read its performance next to its form: low cooling and good daylight, for example, but more heating.
• Strengths and weaknesses. Compare shows up to eight options side by side. It dims the rows where all the options have the same value and marks the highest and lowest value in each row, so the differences stand out.
• New criteria. Looking at alternatives often brings up criteria that are not in the data. You might notice, for example, that the deep-set windows on the south face look right. Record observations like this as a note on the option. Decide explains how to turn one into a dimension of its own.

The example below shows how four common patterns look in the parallel coordinates and in the scatterplot, using the same options for both.

Two dimensions, A and B. An axis without a goal has its low values at the bottom, as drawn here. An axis with a goal has its better values at the top. When both axes have goals, level lines therefore mean that the two goals improve together, and crossing lines mean that they conflict.

Expert mode adds two more aids. Clusters group similar options and show one picture for each group. Analysis reports the spread of every dimension, the outliers, and which inputs affect which outputs.

Look for the alternatives that satisfy your criteria, including criteria that emerge along the way. An alternative is a proposal and not merely a data point, so even an option that performs poorly can lead to a better one. The options you select are not final either. You refine them, or use them to guide the next round of exploration.

Two filters, daylight of at least 56% and cooling of at most 1,850 kWh, leave 210 of the 648 options, and the charts highlight them. The drawer on the left holds a shortlist of three options and the set saved from these filters.

Try it: Open the Windows study · Tour this task · Watch this part

• Filter. Drag along an axis in the parallel coordinates, or drag a box in the scatterplot. Each filter appears as a chip in the bar at the top, next to a count of the options that are left. Remove a chip to remove its filter. Undo and redo step through your filter changes.
• Pick by eye. Drag a box over pictures in the gallery to keep only those options. This lets you narrow by what the designs look like as well as by their numbers.
• Ask. Type a request into the assistant's box in the toolbar, such as cooling under 1850 or what do these have in common. The assistant applies it or answers the question. It runs in your browser. Press / to jump to it.
• Go deeper. Zoom in makes the remaining options the whole collection. The charts rescale to fit them, and you can filter again within them. Exclude does the opposite and removes those options from the collection. Step back undoes the last zoom or exclusion, and Reset restores the full collection.
• Keep proposals aside. Click the bookmark on a card to add that option to the shortlist. The shortlist is not affected by filtering or zooming, so an option that fails a filter but still interests you stays there while you keep looking.
• Save where you are. A set saves the current options together with the filters that produced them. Restore a set to return to that point. You can also keep several sets as alternative directions and compare them later.
• Start the next round. You can load the table you download as a smaller study of its own. The input values of the options you kept also tell you where to rerun the parametric model in finer steps.

Filters in Lite, a selection in Expert

Lite and Expert narrow in different ways. The difference is in how each new brush combines with the ones before it.

Top: three brushes in Lite, with the filter bar after each one. Bottom: the five ways Expert can combine a new brush with the current selection.

In Lite, every brush is a filter, and filters always combine in the same way.

• A second brush on the same axis adds an alternative range. An option passes if it is in either range (OR). A second box in the scatterplot works the same way.
• An option has to pass the filters on every axis (AND), so brushing another axis always leaves fewer options.
• Dragging an existing brush changes its filter. It does not add a new one.

This rule cannot express every combination. In return, your filters always read as a plain list of criteria, and the order in which you made them does not matter.

In Expert there are no filters. You have one selection, and you choose how each new brush changes it. A brush can replace the selection, add to it, subtract from it, toggle it, or intersect with it. Hold Shift to add, Ctrl to subtract, and both to intersect. A step does not have to be a brush. It can also be a cluster, a lasso in the scatterplot, or a single option that you click.

This lets you make selections that filters cannot, such as the options that are bright or cool, except those that need a lot of heating. To get that selection, select the bright options, add the cool ones, then subtract the ones with high heating. Unlike filters, the order of these steps affects the result.

Read where you are

As you narrow, the filter bar at the top shows where you are in the design space.

The filter bar at three points in the Windows study.

• The count shows how many options are left, here 210 of 648.
• Each chip is one criterion. It names the dimension and the range you kept, in the units of the data. A chip such as ≥ 56 comes from a brush that reaches the top of its axis. Click the × on a chip to remove that criterion.
• OR links two brushes on the same axis. In the middle row, the orientation is either about 90 or about 180. Each of the two chips has its own ×.
• A grey chip marks a zoom. In the bottom row, Zoomed in (49) means the collection is now those 49 options, and the count is out of 49. The chip's tooltip lists the criteria that were active when you zoomed. Click its × to zoom back out.
• The name of a set appears next to the count when the current filters are the ones saved in that set. It tells you that you are back at that saved point.

You can read the bar from left to right as a sentence: daylight of at least 56%, and cooling of at most 1,850 kWh, and an orientation of about 90 or about 180. A description like this is easier to remember and to explain to a colleague than a list of 49 option numbers. It is also what DesignSense saves with a set and writes into the note of a captured view.

Expert describes its selection in a similar way. A strip next to the count shows the selection as a single statement that uses AND, OR and NOT. DesignSense simplifies this statement to the shortest one that selects the same options. Click the count to see the full statement and the steps it came from. Undo takes you back through those steps one at a time.

Expert after three steps: select the bright options, add the cool ones, then remove those with high heating. The statement describes the 415 options that are selected.

Combine objective and subjective criteria to make the final choice.

Deciding is difficult. Goals conflict, the people involved weigh them differently, and some of what matters was never measured. No single method resolves all of this, so DesignSense offers several that you can use on their own or together. All of them start from the same step, which is to say what "better" means.

• Compare as you go
  Put options side by side to find out what you like about each one.
• Find the best trade-offs
  Set aside every option that another option beats on all the goals.
• Score the options
  Combine the goals into one number, and try more than one weighting.
• Add your own judgement
  Rate options, write notes, and add criteria that are not in the data.

Try it: Tour this task · Watch this part

Say what better means

Open Dimensions and set a goal for each output, such as lower cooling or higher daylight. Arrows on every axis then show the better direction, and the methods below use the goals to tell which options are better. Columns named with min: or max: already have their goals set when you load them. Open Dimensions

The Dimensions and Goals dialog. It shows the goal for each output, a custom dimension (Looks), a weighted dimension (Energy use) and the star rating.

Compare as you go

Do not save comparison for the end, when you already have your finalists. Comparing two or three options early helps you find out what you like about each of them and which criteria you actually care about. You can then use what you learn to adjust your goals and filters, or to add criteria.

Compare takes up to eight options from the shortlist, the selection or a saved set. It puts their measured values, your own dimensions, your ratings and your notes into one table. Rows where every option has the same value are dimmed, and the highest and lowest value in each row are marked.

The three shortlisted options in Compare. Ratings and notes are at the top, above the measured values, and arrows mark the highest and lowest value in each row.

Find the best trade-offs

The best trade-offs are the options that no other option beats on every goal at once. Together they are called the Pareto front. For any option that is not on the front, there is another option that is at least as good on every goal. You can therefore set it aside without having to decide how much one goal is worth compared with another.

• When to use it. Use it when you cannot, or do not yet want to, put numbers on how the goals trade against each other.
• What it does not do. It does not choose between the options on the front. The more goals you have, the more options end up on the front.
• Where to find it. The scatterplot marks the front for the two dimensions it shows. For the front across all goals, ask the assistant for the best trade-offs, or in Expert use Analysis ▸ Keep the best trade-offs.

Score the options

A weighted dimension combines several criteria into one score. DesignSense scales each measured column from 0 to 1 between its lowest and highest value in the study, and reverses the scale when lower is better. It then multiplies each scaled value by the weight you give that criterion and adds the results. The score becomes an axis like any other. You can sort and colour by it, but you cannot filter by it. An option that is missing a value for any of the criteria gets no score.

The criteria do not all have to be measured. You can also give a weight to a custom dimension that you fill in by hand, such as form, aesthetics or how well the design fits the street, and to your star ratings. Subjective criteria like these then count towards the score with an explicit weight that other people can see and question. Otherwise they tend to influence the decision anyway, but without anyone stating how much. DesignSense scales a custom dimension by the range you gave it (for example 0 to 10), and not by the values in the study. Fill in a value for every option you want scored, because an option with a blank gets no score.

This kind of score is the value function of the Design Space Construction framework (Haymaker et al., 2018). The framework also gives guidance on using it well:

1. Agree on the goals before the weights. Decide who is affected, who makes the decision, and how each goal is measured. The score can only reflect the goals you include.
2. Keep hard requirements out of the score. A limit that must be met is a constraint, so apply it as a filter. Use weights only for goals that can be traded against one another.
3. Base the weights on people's priorities. Each weight is a share of 100%. It states how much that goal should count for the people the decision affects.
4. Try more than one weighting. The paper notes that alternatives "may rank differently depending on the set of preferences used for evaluation". Make a weighted dimension for each stakeholder, or for each view of the priorities, and check whether the same options stay near the top. If they do, the choice is robust. If they do not, you have found a disagreement that the group needs to discuss.
5. Do not rely on the score alone. The score ranks the options but does not tell you why one ranks above another. Inspect the top options in the charts and the gallery to see what trade-off each one makes. Expect to go back and change the goals and weights as you learn more.

The example below scores the same six options under several weightings.

Six made-up options scored on three measured criteria (daylight, cooling and cost) and one judged by hand (looks). The ranking changes with the weights. An option that stays near the top under every weighting is a robust choice.

Weighted scoring is a reasonable method when:

• several goals conflict,
• there are too many options to compare by eye,
• the goals are agreed and every option has a value for each of them, and
• the people involved can say how much each goal matters.

It is less suitable when:

• a poor result on one criterion cannot be made up for by a good result on another,
• only a few options are left, in which case you can compare them directly, or
• people still disagree about the priorities. In that case, use it to explore the disagreement, not to settle it.

Remember that the score is relative. Each measured criterion is scaled against the options you generated, so a high score means good compared with those options, and not necessarily good in absolute terms.

Add your own judgement

Some of what matters will not be in the data. DesignSense gives you three ways to add it:

• A custom dimension holds a value that you enter by hand for each option, for example for buildability, for how well it fits the street, or for the client's reaction.
• Star ratings record how much you like an option. They become an axis that you can filter by.
• A note records your reasons, for example four stars, but check glare on the south face.

You can give custom dimensions and ratings a weight in a weighted dimension, just like measured columns, so that measured and judged criteria count in the same score.

Share your results with others to support the decision.

All of the options below are in the Share menu in the toolbar. Tour this task · Watch this part

The summary image for the shortlist. It shows the shortlisted options, how they compare with the rest, and findings that you can edit before you copy or download it.

Two things to know about the link:

• The link contains only the dataset. Your filters, sets, shortlist, ratings and notes stay in your browser. To share those, send the CSV, the summary image or the report.
• You can only get a link for a study that you opened from a link. A study opened from your computer has no address to share, so put it on GitHub first.

What DesignSense saves. For each study, DesignSense saves your sets, shortlist, ratings, notes, custom and weighted dimensions, goals, column choices and story in your browser. They are restored when you open the same study again in the same browser. Filters are not saved between visits. To delete saved work, use Help ▸ Clear data.

DesignSense was developed through a design study with practitioners in the built environment.

Designing with sense: A critical review and proposal for enhanced design space exploration in generative design
Ahmed M. Abuzuraiq and Halil Erhan · International Journal of Architectural Computing, 2024

DesignSense is a research project and is still changing. We welcome feedback: use Help ▸ Send feedback in the app, or open an issue. The code is released under the MIT licence.