UCB acquisition (κ=1.0) in surrogate Bayesian optimization will outperform EI in predicting room-temperature superconducting phases of hydrogen-rich materials (e.g., H₃S) by prioritizing high-uncertainty pressure regimes (200–500 GPa), analogous to its validated success in drug discovery.

In a closed-loop Bayesian optimization (BO) task that searches hydrogen-rich compound compositions and pressures (50–500 GPa) for maximal DFT/DFPT-predicted superconducting critical temperature (Tc, via McMillan-Allen-Dynes or Migdal-Eliashberg formalism), an acquisition function using Upper Confidence Bound with fixed exploration weight κ=1.0 will identify a candidate structure with Tc within 5% of the best Tc discoverable by exhaustive/grid search, using at least 30% fewer DFT single-point evaluations than Expected Improvement (EI) acquisition, across ≥5 independent runs with different random seeds and initial designs, on at least 2 of 3 benchmark hydride families (H3S-like sulfur hydrides, LaH10-like lanthanide hydrides, YH9/YH6-like yttrium hydrides).

1. Construct 3 benchmark landscapes (H-S, La-H, Y-H systems) using precomputed or newly computed DFT Tc surfaces over a discretized (composition, pressure) grid (~500–2000 grid points per family) as ground truth.
2. Fit surrogate BO loop: GP surrogate + UCB(κ=1.0) vs GP surrogate + EI, both warm-started with identical 10 random initial points per run.
3. Run 5 seeds × 3 families × 2 acquisitions = 30 runs, budget-capped at 150 sequential DFT queries per run (queries resolved by lookup against precomputed ground-truth grid to avoid recomputing DFT every run — critical for tractable cost).
4. Compare: (a) evaluations-to-threshold (reach 95% of global best Tc in grid), (b) best-Tc-found at fixed budgets (25/50/100/150 evals), (c) regret curves.
5. Sensitivity analysis: repeat with κ ∈ {0.5, 1.0, 2.0, 5.0} and EI with ξ ∈ {0, 0.01, 0.1} to test robustness of any observed gap.
6. Validate top-3 UCB-discovered candidates per family with a fresh, independent full DFT relaxation + Eliashberg Tc calculation (not lookup) to confirm surrogate fidelity.

for family in [H-S, La-H, Y-H]:
    grid = load_or_compute_dft_tc_surface(family)   # (composition, pressure) -> Tc
    surrogate_gp = fit_baseline_GP(grid.sample(10))
    assert loo_cv_r2(surrogate_gp, grid) > 0.6       # gate: abort if surrogate can't fit

    for acquisition in [UCB(kappa=1.0), EI(xi=0.01), PI, Random, Grid]:
        for seed in range(5):
            gp = GP(kernel=Matern52())
            X_init = sample_initial_points(grid, n=10, seed=seed)
            gp.fit(X_init, grid.lookup(X_init))
            history = []
            for step in range(150):
                x_next = acquisition.argmax(gp, grid.domain)
                y_next = grid.lookup(x_next)          # simulated "DFT query" via lookup
                gp.update(x_next, y_next)
                history.append((step, x_next, y_next, regret(grid, y_next)))
            log_run(family, acquisition, seed, history)

results = aggregate(logs)
stats = wilcoxon_paired(results, "UCB", "EI", metric="evals_to_95pct_threshold")
sensitivity = sweep_kappa_and_xi(family_subset, kappa=[0.5,1,2,5], xi=[0,0.01,0.1])
top_candidates = select_top3_per_family(results, acquisition="UCB")
independent_dft_validation(top_candidates)  # real VASP+EPW run, not lookup