# Quantum Wave — first neural training pilot

Recorded 23 September 2026

## Outcome

A real FermiNet H2 training run completed 1000 VMC optimization steps on one NVIDIA RTX 4090
The final checkpoint was restored successfully and all model parameters were finite
A second fresh training run with the same configuration and seed completed on the same machine
These runs establish an engineering training workflow and internal numerical repeatability
They do not establish independent scientific validation or a converged ground-state energy

## Configuration

- Neutral singlet H2 with a bond length of 1·4 bohr
- Continuous-space all-electron Coulomb Hamiltonian including nuclear repulsion
- Four determinants and two hidden layers with widths (64,16)
- 256 walkers and 10 Metropolis steps per optimization iteration
- 100 Hartree–Fock pretraining iterations and 200 MCMC burn-in iterations
- Adam optimizer with initial learning rate 0·001 and the recorded upstream decay schedule
- Training seed 23 and 64-bit JAX arithmetic
- 17680 model parameters

## Actual execution

| Measurement | First run | Same-environment repeat |
|---|---:|---:|
| Completed training steps | 1000 | 1000 |
| Training call wall time in seconds | 47.6601 | 45.1141 |
| First logged training energy in hartree | -1.0153333738 | -1.0153333738 |
| Last logged training energy in hartree | -1.1701313977 | -1.1701313977 |
| Last 100 training steps mean in hartree | -1.1737814435 | -1.1737814435 |

Runtime covers the training function including its pretraining, burn-in and compilation work rather than installation or artifact preparation
The listed energies come from the changing model during training
No independent uncertainty estimate is attached to these numbers
The last-100 mean must not be presented as the final independently evaluated energy

## Internal repeatability

Maximum absolute difference across the two training energy traces was 4·885e-15 hartree
Maximum absolute difference across exported model parameters was 6·523e-15
Both were below the internal numerical comparison tolerance of 1e-12
The same hardware, software, seed and operator were used
This is not a fresh-seed evaluation or an independent researcher reproduction

## Preserved artifacts

The execution archive contains the resolved configuration, training log, every-step CSV, final upstream checkpoint, numeric-only parameter export, environment record and source revision references
SHA256 checksums verify the copied files
The upstream checkpoint is a pickle-capable format and should only be loaded when its source is trusted
The separate numeric parameter archive loads with pickle disabled and was round-trip checked against the restored checkpoint
No website service executes these files

## Tested environment

Linux x86_64 and NVIDIA driver 595·71·05
Python 3·12·14, JAX and JAXlib 0·8·2 with CUDA 12 runtime packages
FermiNet commit c4312c315dda1c5728994ba89629744f71c6eb66
KFAC-JAX commit 086619d3b44749e194fa5721c4a3bdb9eb8e3b2c
folx commit eec79b6871502b075b7adaf70c96abddbcfe4b95
Exact package versions are retained in the dependency lock

The final validated optimizer is Adam
The initial KFAC configuration failed dependency compatibility checks and is not the claimed completed run
No upstream scientific algorithm source was modified
A wrapper ensures interval and final checkpoint saves

## Next stage

Implement a fixed-parameter evaluator with explicit fresh seeds, new walkers, fresh burn-in and autocorrelation-aware uncertainty estimates
Then ask a separately authorized researcher to execute the frozen evaluation protocol in another environment
See the accompanying stage-two plan for implementation and acceptance gates