Experiment Platform: ThermCircuitEnv

All experiments in this research program share the same core loop: configure a Langevin particle system with an energy landscape, wire it into a training environment, run a gradient-free optimizer or RL algorithm to modulate temperature or other control channels, and measure the result. The ThermCircuitEnv crate eliminates the boilerplate that each experiment would otherwise repeat.

What It Provides

ThermCircuitEnv sits on top of two layers of infrastructure:

The ML chassis (sim-ml-chassis) provides the Algorithm trait, batched environments, policy representations, and rollout machinery. Every algorithm (CEM, REINFORCE, PPO, TD3, SAC from sim-rl; SA, Richer-SA, PT from sim-opt) bolts onto the same trait. Any experiment built on ThermCircuitEnv automatically gets access to all 8 algorithms with zero additional wiring.

The thermostat layer (sim-thermostat) provides LangevinThermostat and passive energy landscape components: double-well potentials, oscillating fields, pairwise coupling, ratchets, and external fields.

ThermCircuitEnv bridges these: it takes a circuit description (particle count, energy landscape, control channels) and produces a ready-to-train environment.

Example

#![allow(unused)]
fn main() {
ThermCircuitEnv::builder(n_particles)
    .gamma(10.0).k_b_t(1.0).timestep(0.001)
    .with(DoubleWellPotential::new(3.0, 1.0, 0))
    .with_ctrl_temperature()
    .reward(|_m, d| -(d.qpos[0] - 1.0).powi(2))
    .sub_steps(100).episode_steps(1000)
    .build()
}

Validation

The platform was validated with a CEM training test: one particle in a double well with ctrl-temperature. CEM learned state-dependent temperature control, achieving roughly 2x improvement over a constant-temperature baseline. This confirms the plumbing is correct end-to-end but does not test the scientific claims of any specific biological regime.

Location: sim/L0/therm-env/ — 38 tests total (31 debug, 7 require --release).