Bella

Discovery in minutes, not months.

Bella screens thousands of materials and protein candidates, filters through foam mechanics in microseconds, and delivers a confirmed shortlist. Bob finds the candidates. Foam mechanics eliminates 99%. The rest get simulated.

2,944
Phonon-stable materials found
53,349
Protein candidates screened in 6.2 min
16
Simulation domains

What Bella does

Materials pipeline

Discover, screen, confirm. One command.

Protein screening

53,349 candidates across 995 disease targets. 6.2 minutes.

Generative mode

No database match? Bella generates candidates from scratch.

16 simulation domains

Nitrogen fixation to fusion plasma. One flag to switch.

Mathematics

Riemann zeros to 10^8. Collatz. Custom PDE/ODE solver.

Technosignature pipeline

6,787 stellar candidates from 28.9M Breakthrough Listen hits.

Bob

Bob is the search layer. As one example run: queried GNoME (554,000 structures), Materials Project, NIST, and Catalysis-hub simultaneously, returning 2,944 phonon-stable materials across 96 application categories. Plug in any materials database. Bob handles the rest.

Query Type Speedup
Wide query 346x
Tight query 1313x
Batch x1000 101x

Index build: 0.64s (one-time) - AVX2+ compatible, AVX-512 used automatically

Foam mechanics

The Kun Framework derives structural validity from first principles. Before any simulation runs, foam mechanics scores every candidate by whether its geometry can exist on the surface of what is physically possible.

The numbers:

Screening cost: microseconds per candidate. No DFT, no GPU.
Protein screening: 53,349 candidates screened in 6.2 minutes.
Phonon accuracy: 89% on 28-material validated set.
QCD calculator (T32): 1.71ms vs lattice QCD hours.

A phonon stability run takes 20-30 minutes per material. Foam mechanics scores the same candidate in under a millisecond at 89% accuracy. Run 5,000 candidates through foam first. Simulate the 50 that survive.

99% eliminated before the first simulation runs.

NSMace

Bella uses NSMace for force evaluation. No separate install.

C++ MACE force field engine. No Python, no PyTorch, no GPU. Runs energy and force calculations on any CPU. Scales linearly with atom count.

System size NSMace Python MACE (CPU)
5 atoms 39.8 ms 400.9 ms
100 atoms 143.6 ms timeout (>5 min)

The funnel

5,000
candidates
Bob + Bella
foam filter
seconds
500
survivors
NSMace
energy + forces
seconds
300
confirmed
SPARC
quantum screen
~5 min
Top 5
shortlist
Phonon
stability runs
~30 min
Final shortlist

Foam mechanics filters ~99% of candidates before any quantum calculation runs. What used to require months on GPU clusters takes under an hour on a laptop.

Requirements

CPU AVX2 or newer (Intel Haswell 2013+ / AMD Zen 2020+). AVX-512 used automatically when available.
RAM 16 GB recommended
Python 3.8+
OS Linux native. macOS native. Windows via WSL.
SPARC Optional. Linux only. Not required for most runs.
GPU Not required
License Simulation →