Hyperparameter Search Farm

Parallel search over model architectures and training settings that tunes Kronos PINNs, GNNs, and surrogates at scale.

Tuning as a first-class workload

A PINN's loss weighting, a GNN's depth, a surrogate's architecture, these hyperparameters strongly affect model quality, and the right values are not known in advance. Kronos runs a hyperparameter search farm on L0: many training runs launched in parallel across the search space, each an independent job, evaluated and compared to find configurations that generalize.

An embarrassingly parallel search

Like the neutronics and shape sweeps, hyperparameter search is embarrassingly parallel: each trial is independent, so the farm scales out cleanly on elastic cloud GPUs and releases capacity when the search ends. Trials are checkpointable and preemptible, making them ideal backfill for the scheduler between larger reserved jobs.

• Search over architecture, depth, and width
• Loss-term weighting for PINNs
• Learning-rate schedules and regularization
• Early stopping and pruning of weak trials

Efficient search strategies

Brute-force grid search wastes compute, so Kronos uses smarter strategies: Bayesian optimization that models the objective and samples promising regions, and early pruning that kills clearly losing trials before they finish. This concentrates GPU time on configurations likely to matter, cutting the cost of a thorough search substantially.

The search matters most for the physics-informed models. A PINN with mis-weighted loss terms can satisfy data while violating the governing equation, or vice versa; the search finds the balance that respects both. For the breeder equilibrium PINN and the burner ambipolar surrogate alike, tuning is what turns a plausible model into a trustworthy one.

Search results feed the retraining pipeline: the chosen configuration becomes the recipe retraining uses each cycle, and searches are re-run when the data or requirements shift. The farm is how Kronos keeps its model architectures matched to the physics they must capture.