Abhinav (@Abhinavsns) on X

Here Plasma starts as hollow Deuterium-Helium-3 (D-³He) ion rings with vortex flow and spin, moving through long tapered formation tubes into the narrow compression chamber. (2/n)

The Resolution? ~ 750K particles with 56×56×112 grid/field lattice. This GPU run took 44.6 seconds on my 14-core Metal GPU in my M3 Pro. No cluster. Developing locally allows rapid turn around times. (3/n)

We can see the FRCs travel, merge and compress, then inspect confinement, vorticity and magnetic structure. Change the tube shape or pulse timing and run it again. (4/n)

This is not full Trenta physics or a fusion-yield prediction as the magnetic wells and compression are prescribed but the full Maxwell evolution and wall physics with calibrated D-He3 reactions can be added easily. (5/n)

Comparatively,

's TORAX is a 1D radial tokamak core-transport simulator that combines RL or AlphaEvolve for real-time control. 3D systems usually represent the geometry and boundary physics accurately for newer asymmetric systems like Helion. (6/n)

The Caveat? 3D is slow. But not anymore with GPU native codes. Local execution matters for physical AI because the loop must stay tight and fast for on the fly reinforcement learning: Measure, Simulate and choose the next control change. (7/n)

You do not want every pulse or robotic movement simulation waiting on a cloud queue before action. Or proprietary reactor/robotic data leaving the machine. Edge AI is the future. (8/n)

Imagine reinforcement learning running thousands of variants locally with agents in the loop: change the geometry, coils or pulse timing, score the result, keep what works. Endless possibilities (9/n)