Tests whether real-world water chemistry — the ionic composition of feed water in green hydrogen electrolysis — measurably alters the optimisation dynamics of machine-learning models used to control such systems. Amended 25 August 2026. As published this was framed as "leveraging validated LMC scaling" to bridge hydrogen systems with precision-induced AI barriers. That scaling law is retracted, and it was never load-bearing here: the question of whether water chemistry affects model optimisation stands on its own and does not depend on any claim about floating-point precision regimes.

Feed-water ionic composition (specifically conductivity, hardness [Ca²⁺/Mg²⁺], chloride concentration, and pH buffering capacity) measurably alters the convergence dynamics, sample efficiency, and final control performance of machine-learning models (e.g., RL controllers, surrogate optimizers, MPC-with-learned-dynamics) used for real-time optimization of PEM/alkaline electrolyzer-based green hydrogen systems, relative to models trained/optimized under idealized (deionized or synthetic-constant-composition) water chemistry assumptions. Falsifiable form: across ≥5 distinct feed-water chemistries spanning realistic industrial/regional ranges (conductivity 50–5,000 µS/cm; hardness 0–500 mg/L CaCO₃; Cl⁻ 0–2,000 mg/L), a model class trained under idealized water chemistry will show a statistically significant (p<0.05, paired/blocked design) degradation in ≥1 of: (a) episodes/iterations to convergence, (b) steady-state control cost/regret, (c) generalization error when transferred to an unseen chemistry, compared to a model trained/adapted with realistic chemistry variation included.

Minimum viable test: a factorial simulation study (no physical electrolyzer required for MVP) using an open-source or custom PEM electrolyzer degradation surrogate (e.g., adapted from NREL H2A/H2FAST degradation curves or literature Nernst-Butler-Volmer models with ionic-strength-dependent membrane resistance term), controlled by 2 model classes (a model-free RL agent, e.g., PPO/SAC, and a learned-dynamics MPC), under 6 water chemistry regimes (1 idealized baseline + 5 real-world profiles drawn from published regional water quality datasets — e.g., USGS/EU water framework directive tap-water surveys). Each condition run with n=10 random seeds. Primary outcome: training curves (steps-to-convergence), final control cost (kg H2/kWh regret vs. theoretical optimum), and cross-chemistry transfer error matrix (5x5).

# Phase 0: Surrogate construction
build_electrolyzer_surrogate(
    base_model="Butler-Volmer + Nernst membrane resistance",
    ionic_dependence_terms=["conductivity", "Ca_Mg_hardness", "Cl_concentration"],
    calibration_target=literature_stack_curves
)  # target R^2 > 0.85

# Phase 1: Environment
class ElectrolyzerEnv(gym.Env):
    def __init__(self, water_chemistry_profile):
        self.surrogate = surrogate
        self.chemistry = water_chemistry_profile
    def step(self, action):
        state = self.surrogate.update(action, self.chemistry)
        reward = f(h2_output, efficiency, degradation_penalty)
        return state, reward, done, info

# Phase 2: Training matrix
for model_class in [SAC, LearnedMPC]:
    for chem_regime in [idealized, real_1, ..., real_5]:
        for seed in range(10):
            train(model_class, ElectrolyzerEnv(chem_regime), seed)
            log(convergence_steps, final_regret, degradation_trace)

# Phase 3: Transfer test
for source in chem_regimes:
    for target in chem_regimes:
        transfer_error[source][target] = evaluate(model[source], Env(target))

# Phase 4: Stats
mixed_effects_model(endpoint ~ chemistry_regime + (1|seed), correction="holm")
ablation: chemistry_aware_obs vs chemistry_blind_obs