Interactive Stochastic Thermodynamics of Neural Network Training
ThermoTrainLab is an interactive web simulator exploring the deep physics connecting stochastic gradient descent (SGD), free energy minimization, entropy production, and Landauer's principle in neural network training.
Built in response to the Bay Area thermodynamic computing revolution — Extropic's 10,000× efficiency claims and Normal Computing's $50M Samsung raise — this tool makes the physics of AI training tangible and interactive.
Watch 50 SGD particles navigate a loss landscape while the free energy decomposition F = U − TS updates in real-time. Adjust learning rate (temperature) and batch size (noise scale) to see how they affect exploration vs exploitation.
Visualize the irreversibility of gradient descent. Track total entropy production, system/environment entropy changes, work done, and the Jarzynski equality connecting non-equilibrium training to equilibrium free energy.
Every bit erased during a weight update costs at least kT·ln(2) ≈ 2.87×10⁻²¹ J. See how modern GPUs operate at 10⁶−10⁹× above this fundamental limit, and why thermodynamic computing matters.
A live Ising model simulation showing how probabilistic bits (p-bits) — the core of Extropic's thermodynamic chips — naturally sample from Boltzmann distributions. Watch the phase transition at the critical temperature T_c ≈ 2.27.
Compare GPU and thermodynamic chip energy consumption across model sizes on a log scale. See the potential 100-10,000× energy savings for probabilistic AI workloads.
# Just open the HTML file
open index.html
# Or serve locally
npx serve .- SGD as Free Energy Minimization (2025)
- Stochastic Thermodynamics of Associative Memory (2026)
- Extropic: Thermodynamic Computing From Zero to One
- Thermodynamic Bounds on Energy in DNNs (2025)
- Normal Computing Raises $50M from Samsung
Single-file HTML + vanilla JavaScript + Canvas API. Zero dependencies. Zero build step.
MIT — Built by Surya Midde as Daily Webapp #18