AI Spacecraft Design: Lighter Parts with a Hidden Fail-Safe Flaw | Om Mali

Software can now design spacecraft parts that are lighter and stiffer than anything a human engineer would draw. In this episode of AI in Astronomy, Om Mali looks at AI shaping the hardware we send into space, and at the one thing it quietly throws away while doing so. After ten episodes about AI reading the sky, this one covers AI building what goes up into it.

How generative design works

You give the software a volume, the loads, and a mass budget. It returns a shape no engineer would draw. It gets there by deleting everything that is not carrying your load. The video points out that this is not just a demo. The bracket holding the S-band antenna on Sentinel-1 was designed this way, going from 1.4 kilograms down to 940 grams, about a third lighter on a real satellite.

Rebuilding the problem

To test the idea, Om Mali built the same kind of problem himself: a block of metal the optimizer may carve, a bolted root, one lug, one load, and a budget to use only 40 percent of the metal. Material flows toward whatever carries load and starves everywhere else. Nobody drew the result, and it answers the question that was typed. Compared with a plate machined down to the same 40 percent, the optimized bracket is 23 percent stiffer at the same mass. The video says plainly that the pitch is true.

The hidden flaw: no second path

Looking at where the load travels, every member of the optimized bracket is working. There is no second path, because a second path would have been spare mass, and spare mass is what the optimizer was told to delete.

The test is to drill one small void in the worst place, less than 1 percent of the area. The plate becomes 20 percent softer. The bracket becomes 31 times softer, from the same void. The video notes this is a known problem: the fail-safe literature says that optimizing for one load case yields a structure with no alternative load paths, and aerospace requires them. Nobody asked the optimizer for them.

Buying back the margin

Giving the optimizer 60 percent of the mass instead of 40 percent changes the picture. The worst void then costs 60 percent of stiffness instead of a factor of 31, and the bracket is still stiffer than the plate. The cost is that the mass saving falls from 60 percent to 40.

The fix is a better brief

The video's conclusion is that the optimizer does what you wrote down. Lighter was in the brief. Surviving damage was not. The fix is not a better optimizer but a better brief.

Key takeaways

• Generative design returns shapes no engineer would draw, lighter and stiffer, by deleting material that does not carry load.
• The Sentinel-1 S-band antenna bracket was designed this way, from 1.4 kg down to 940 g.
• In the experiment, the optimized bracket was 23 percent stiffer than a plate at the same 40 percent mass.
• A void under 1 percent of the area softened the plate by 20 percent but the optimized bracket by a factor of 31.
• Optimizing for a single load case leaves no alternative load paths, which aerospace requires.
• Allowing 60 percent mass reduced the damage penalty to 60 percent, at the cost of a smaller weight saving.
• The optimizer only optimizes what you wrote down.

Try it yourself

Take something you have already shipped and ask your AI model: here is what I asked you to optimize; tell me what I left out of the objective and what that omission would cost in the real world. The video ends with the idea that the missing requirement, not the tool, is where the risk lives.