Rakesh
Balamurugan

PhD candidate, UConn

I build computational methods and AI to design complex engineering systems and predict how they fail.

Featured Research

A-kd-tree

How far apart are two objects, out of 1.67 million spheres? Watch the search answer it, then see the time against the state of the art on GPU and on CPU.

Time to answer one frame

GPU0 ms

ours

…

SOTA

…

CPU0 ms

ours

…

SOTA

…

1 · The question

Two objects in a crowded scene. How far apart are they?

Fatigue life of 3D-printed metal

How many load cycles will a printed titanium part survive? Two methods predict it from the CT scan: a phase-field simulation and a neural network.

1 · Loaded again and again

A propeller blade is stressed and relaxed with every turn. Over millions of turns, that can crack it.

The settling motion and the break are illustrations. Methods are under review.

More research

Rays travelling from one triangle-mesh bunny to another, each stopping where it hits the second surface

RT-accelerated geometric queries

24×

faster than the A-kd-tree CUDA query kernel

Ray-tracing hardware used as a search engine for exact distance queries between objects.

CUDANVIDIA OptiX

Burgers shock: the numerical-solver field next to the error maps of the state-of-the-art neural operator and ours

Neural operators that generalize like numerical solvers

16%

less error than the state-of-the-art neural operator, same parameter count

State-of-the-art neural operators smear sharp fronts. Ours stays closer to the numerical solver, right where the shock is.

PyTorch