Project DTM

Denoising Thermodynamic Models — can an energy-based model on a thermodynamic sampler actually be trained?

The full program in three pillars: the theory, every experiment, and the stack. Scroll, or read the notebook.

Theory

One quantity decides whether a model learns — and the textbook proxy for it is wrong by ~10³⁰×.

The Q gradient-SNR predictor, the spectral-gap symmetry zero, and the observable-projected fix that tracks 45 of 48 exact-diagonalization cells.

Experiments

Every claim tested against a frozen pre-commitment, exp1 through exp19.

Exact diagonalization, block-Gibbs RBMs, and GPU-scale DTM-MNIST runs — all logged and version-tracked on Weights & Biases.

Stack

The exact toolchain that turns a claim into a result anyone can re-run.

Python · JAX · thrml · Equinox · Weights & Biases · rented H200 — pinned versions, git-frozen pre-registration, measure-only discipline.