Design formulations from scratch
Tell Backbond which properties to move and hand it the ingredients you can actually buy. It designs candidate recipes, grades every one against every target, works out where those targets fight each other, and hands back the few worth making.
A closed loop, not a one-off model
Every run is designed against the graph your own results built, and every result you bring back sharpens the next one. The search gets better at your chemistry because it is reading your chemistry.
You state the problem
The product, the properties that have to move and which way, the ingredients you can actually source, and how many experiments the bench can run. That is the whole brief.
It searches the space
Hundreds of candidate recipes per run, each one composed against your graph and graded on every target at once, with the reasoning behind each grade kept.
You get a shortlist, not a ranking
The few experiments chosen to cover the most ground on the bench, which is a different list from the top scores, and the trade-offs that made it that list.
State the problem, not the method
What the product is, which properties have to move and in which direction, the ingredients available to you, and how many experiments you can afford. No objective function to write, no weights to tune, no solver to configure.
- Targets carry a direction: raise, lower, or hold where it is
- Ingredients are yours, so nothing comes back that you cannot source
- Start from a saved baseline recipe, or from nothing at all
Hundreds of candidates, graded on every target
The engine composes candidates against the graph, predicts every target for each one, and keeps the reasoning behind each prediction. It streams while it runs, so the design space fills in as the search explores it rather than appearing at the end.
The whole search, on one map
Every candidate projected into the space its own composition defines, so recipes that are alike sit near each other and the shortlist can be checked for spread. The promise of each candidate is its shading, so the good regions are visible as regions rather than as a list of rows.
- Rotate into three dimensions when two are not enough
- Click any point for its full composition and predicted properties
- Clusters are real: near on the map means near in the recipe
The shortlist, target by target
Each chosen experiment against each target, on its own axis, so a recipe that wins on three targets and loses on the fourth is legible at a glance. This is the view that turns a search result into a decision about what to put on the bench.
Where two targets fight each other
- A frontier is the set of recipes where you cannot improve one target without giving up another. Everything behind it is dominated and can be dropped.
- The slope between two points is the exchange rate: exactly what a point of fracture toughness costs you in tensile modulus, in the units you work in.
- The shortlist is picked along the frontier rather than from the top of a ranking, so the experiments you run bracket the decision instead of crowding one corner of it.
- Every point stays clickable back to its full composition, so a trade-off is never an abstraction.
Frequently asked questions
How much data do I need before this is useful?
Less than you would expect, and that is the research problem Backbond is built around. The search runs over the causal graph your own documents produced, so it starts from what is already known about your ingredients rather than from scratch, and it optimizes in the regime where there are hundreds of results rather than millions.
Can I constrain it to ingredients I can actually buy?
Yes, and that is the normal way to run it. The brief takes the ingredient list you can source, and the search only composes candidates from it. You can also hold a baseline recipe fixed and ask for variations around it.
What is the design space map showing?
Every candidate the run produced, projected into the space its own composition defines, so recipes that are alike sit near each other. It is how you see whether a shortlist is spread across the space or clustered in one corner of it.
Why a Pareto frontier and not a single best recipe?
Because there usually is not one. When two targets pull against each other there is a set of recipes where you cannot improve one without giving up the other, and the useful question is which point on that edge you want. Naming a single winner would be hiding the decision rather than making it.
What do I do with the shortlist?
Run it. The experiments come back into the same project, the graph learns from the results, and the next run starts from a better model than the last one did. That loop is the product.
Bring one problem, and four short lists
The search, the modelling and the shortlist are on us. Bring a project, the ingredients you can source and the properties you have to hit, and we will run it with you.