Show HN: Physics-based CAD | Hacker News Reader

Show HN: Physics-based CAD

prandtl.design

104 pointslochieferrier27 comments

This is not Physics-based at all, you need to introduce other variables like wind speed around the drone, the direction of the wind, angle of attack, viscosity, etc. This example is parametric modeling at best or a configurator, not saying that is bad, just has a misleading title. I do not see any real geometry manipulation or groundbreaking CAD being attempted. Here is a viewer that Autodesk released this year using javascript.

https://developer.autodesk.com/

For sure, there's a lot missing, title is probably misleading. Definitely nothing new in terms of the CAD, browser CAD has been well done, just look at professional products like Onshape. Right now it's primitives only for this, one of the next steps is to look at how full meshes can be optimized. A simple example of this could be wings: https://people.eecs.berkeley.edu/~pabbeel/papers/2013_gp_win...

It's called parametric modelling, and many CAD software use a parametric kernel (solver). https://en.wikipedia.org/wiki/Computer-aided_design

Is this somehow different? Please explain

The difference is that there is a simple physics model that's driving the dimensions. So the arms are getting resized according to beam bending, the battery is a certain size for the energy requirement, motors are sized by kw/kg and kw/m^3 and so on.

There's definitely ways to make this sort of model in something like Solidworks through the use of a variables interface that interacts with the drawing dimensions. Though it can't do any sort of fast optimization with the physics relations. What's happening here is a guaranteed global optimal design is being returned for a given set of inputs and objectives in a few seconds.

This sounds interesting but I can't react the site. Could you please explain how you get a guaranteed optimum? What if you have conflicting objectives?

I think by react you meant react to, which was fair enough, it was dodgy as. Now running well.

The property of getting a guaranteed optimum is something that arises from setting the design problem up as a geometric program. You won't always get a solution (for example, a certain size quad has a minimum prop diameter). Though if you do, it is guaranteed to be optimum because the problem is convex as outlined here: https://gpkit.readthedocs.io/en/latest/gp101.html

Also, conflicting objectives are naturally traded off until you find a Pareto optimal solution. Using geometric programs lets you balance conflicts in a design quickly. This is why they are super useful for aerospace applications, where you are always looking at a tradeoff between weight and at least one other variable.

I meant to write reach, sorry. Glad that it's working well now. Thanks for the reply, and for the link, it sounds like something fun to get into =).

It forces you to assign a weight to each objective, converting from a multi-objective to a single-objective optimization problem (the weighted sum method).

By varying the weights and re-running, you could identify a set a solutions on [convex regions of] the Pareto front.

This is correct, much better than what I said.

While this is well made, in production you would probably put the physics model in Excel with a bunch of VB scripting, and interface it to parameters and equations in solidworks. This isn't something I've done, but I absolutely wouldn't want to trade in a cad engine that can do everything badly just for some design tool that does a few things well.

if you want to get somewhere with this, you should consider the engineering work flow. usually there is a conceptual/design phase that consists of a lot of photographed whiteboards, physical mockups, Matlab, even Mathematica Notebooks if you're like that. Or actually, some cfd and multiphysics fem modelling if you are serious about your flying objects.

You seem to fit into this chaos sonewhere, but I wouldn't quite know at which point I should use this. After all, cad models are always parametric until you start adding all the details, and actual physics modelling even on a xflr5 level goes well beyond this.

Nice project. What are you hoping to add in the future?

Our team works on goal directed design as well, but at the moment it's less of an interactive optimization tool and more of an involved job running process a la Optistruct. One of our fun projects was a drone body as well :)

link for the interested: https://autodeskresearch.com/projects/dreamcatcher

Wow, that is really beautiful. Topology optimization is cool stuff.

I think it's key for optimization to be fast so that the designer can get a feedback loop going between the computer and the ideas in their head. Would love to talk about whether geometric programs might help Autodesk generate topology faster.

Near future stuff is probably a real-world component input (optimized shopping), and making things more flexible so you can drag and drop something together from components.

Cool, I think more applications should be using solvers like this.

What solver are you using? Is it solved server side?

Currently solved server side, though assembled on the client side (see console) with gpkitjs and then passed through to GPkit.

Packages here: https://github.com/lochieferrier/gpkitjs and here: https://github.com/hoburg/gpkit

Geometric programs are super cool.

Interesting, do you have some more info about how it works? When I make the battery larger, is it considering the weight of the battery? Weight of motors, props, etc?

Yes it's considering the weight of the battery, as well as the weight of the motors. It doesn't factor in the props, I didn't get around to that. The props have inaccurate geometry of course, which I'm looking to fix with a mesher.

Have a look at the console for how the geometric program is assembled.

Cool. I think this tool has a lot of potential users, especially if you can make it pick between commonly available props, esc, motors and so on. Imagine how cool this would be if it could make a list of those things and export a 3d-printable frame :)

I managed to break the tool by inputting some infeasible inputs and then changing them again several times, and then refreshing the page.

This is my first time putting something up, everything is breaking, fixing it now.

No test like a production test.

this is more a configurator than a CAD.

That page is broken

Slashdot Effect probs

Nope, I just have no idea what I'm doing.

hugs

Takes some balls to go live with a project when you're on a step learning curve. Congrats :)

Fixed it, didn't know about pythonpath with gunicorn.

That will pass soon enough.

Reply on news.ycombinator.com