Warm-restart training in surrogate Bayesian optimization (UCB acquisition) escapes local minima in hydrogen sulfide electronic structure calculations, reducing the pressure required to achieve metallization by ≥20% in first-principles simulations compared to fixed-learning-rate baselines.

This discovery has a Claude-generated validation package with a full experimental design.

Precise Hypothesis

Spine & Adversarial ReadReady for validation

“This hypothesis tests whether adding scheduled warm-restarts to a UCB-acquisition Bayesian optimization loop, versus a fixed-learning-rate baseline, reduces the surrogate-predicted metallization pressure of hydrogen sulfide by at least 20% under matched computational budget.”

• highThe claimed effect may simply reflect increased effective exploration (more diverse samples) rather than any specific mechanism of 'escaping local minima' — a fixed-LR baseline given equivalently randomized restarts or a larger initial sampling budget might match warm-restart performance, undermining the mechanistic claim.
  Protocol partially addresses this via compute-matching (≤1.1x core-hours) and trajectory/basin diagnostics (step 10), but does not yet include a third control arm (fixed-LR + random restarts without the UCB-scheduled structure) that would cleanly isolate 'warm restart mechanism' from 'more randomized exploration.' This is a gap: the EVP should add that control arm before claims of mechanism (not just outcome) can be defended.
• mediumWhy UCB acquisition and this specific warm-restart schedule (every 15 iterations, log-uniform perturbation 0.5-2x) rather than Expected Improvement, Thompson Sampling, or a cosine-annealing restart schedule (SGDR)? The methodology choice appears arbitrary and unjustified relative to alternatives well-established in the BO/ML literature.
  Not resolved in current EVP. UCB is chosen presumably for its explicit exploration-exploitation tunability (kappa parameter) which pairs naturally with restart-based re-exploration, but no ablation against EI/Thompson Sampling or alternative restart schedules is included. A sensitivity analysis (step 11) covers restart frequency/magnitude but not acquisition-function choice itself — this should be added as a required ablation, not left as an implicit design decision.
• mediumThe definition of 'metallization pressure' as band-gap crossing ≤0.05 eV threshold via a surrogate model is itself a coarse and potentially unreliable proxy for true metallization (which in real H3S involves subtleties like electron-phonon coupling and superconducting gap formation, not just DFT band-gap closure) — the entire experiment could be optimizing a metric that is disconnected from physically meaningful metallization.
  Partially addressed via full-fidelity DFT cross-validation (step 7, success criterion of ±10 GPa agreement), which anchors the surrogate metric to actual DFT band structure. However, this does not validate against experimental metallization pressure or electron-phonon/superconductivity calculations, so the result would speak to 'DFT band-gap-defined metallization' rather than the full physical phenomenon — this scope limitation should be stated explicitly in any resulting publication.

Experimental Protocol

ROI Projection

Implementation Sketch

# Pseudocode: Warm-restart UCB-BO vs fixed-LR baseline for H2S/H3S metallization search

initialize structure_pool = load_polymorphs(H2S, H3S)  # 6 structures
initialize shared_surrogate = train_GP_or_NNP(initial_DFT_dataset)

for seed in range(5):
    set_random_seed(seed)

    # Baseline arm
    baseline_opt = BOOptimizer(acquisition="UCB", lr_schedule="fixed", surrogate=copy(shared_surrogate))
    baseline_result = run_optimization(baseline_opt, structure_pool, pressure_grid,
                                        convergence_tol=1.0_GPa, max_iters=200)

    # Warm-restart arm
    wr_opt = BOOptimizer(acquisition="UCB", lr_schedule="warm_restart",
                          restart_every=15, perturb_scale=loguniform(0.5,2),
                          surrogate=copy(shared_surrogate))
    wr_result = run_optimization(wr_opt, structure_pool, pressure_grid,
                                  convergence_tol=1.0_GPa, max_iters=200)

    log(seed, baseline_result.P_metal, wr_result.P_metal,
        baseline_result.core_hours, wr_result.core_hours)

# Post-processing
compute_percent_reduction(baseline_results, wr_results)
run_paired_significance_test()
validate_top_candidates_with_full_DFT(wr_result.best_structures)
plot_trajectory_diagnostics(baseline_opt.history, wr_opt.history)  # PCA/UMAP basin visualization