Accelerated design of proton exchange membranes for green hydrogen production with artificial intelligence

Abstract: Water electrolysis is an eco-friendly method for hydrogen production that has reached significant levels of technological maturity. Among commercialized water-electrolysis technologies, proton-exchange membrane electrolyzers offer high current density, fast dynamic response, and compact system design, among other advantages. On the other hand, managing their high capital cost and the ``forever-chemistry'' nature of Nafion, a perfluorinated proton-exchange membrane widely used in such devices, remains a major challenge. Searches for fluorine-free replacements for Nafion, pursued largely through physical experimentation, have been active for decades with limited success. In this work, we develop and demonstrate an AI-based strategy for designing new proton-exchange membranes for electrolyzers. Two key components of this strategy are an implementation of the virtual forward-synthesis approach and a set of machine-learning predictive models for essential application-inspired membrane properties; the former generates a vast space of millions of synthesizable polymers, which are then evaluated and screened by the latter. The strategy is validated against experimental data for known membranes and then applied to design over 1,700 new synthesizable candidates. This article concludes with a forward-looking vision in which the strategy could be elevated into an interactive and iterative scheme that are based on LLMs to facilitate materials design in multiple ways.

• The paper presents an AI-driven method that virtually synthesizes and screens 66 million polymers to identify 1,738 halogen-free, high-performance proton exchange membrane (PEM) candidates for green hydrogen production, all of which match or exceed Nafion's key properties under required operating conditions.
• The study's screening identifies sulfonated aromatic polyimides as the most promising candidates for optimized PEMs, thanks to their rigid structural features and performance enhancing synergy.
• The machine learning models exhibit excellent accuracy, with R2 values consistently above 0.82 for relevant properties, validating the method's initial success.

Motivation and problem statement

Proton-exchange membrane (PEM) water electrolyzers are a commercialized route to green hydrogen, offering high current density, fast dynamic response, high-pressure operation, and compact system design. Their principal drawbacks are capital cost and the incumbent membrane: Nafion, a perfluorinated sulfonic acid copolymer whose exceptional proton conductivity and durability stem from extremely strong C–F bonds that also make it effectively non-recyclable — the authors characterize it as "forever chemistry." Decades of experimentally driven searches for fluorine-free replacements (e.g., Pemion and sulfonated poly(2,6-dimethyl-1,4-phenylene oxide), sPPO) have yielded only partial successes, with conductivity, stability, or durability still trailing Nafion. The paper's premise is that the polymer design space is too large for serial experimentation and that a virtual forward synthesis (VFS) plus ML screening workflow can compress the search.

Design criteria

The authors translate electrolyzer operating requirements into eleven quantitative criteria. The most consequential is proton conductivity $\sigma > 0.1$ S/cm at 80 °C and 100% relative humidity (RH), matched to Nafion's performance. Notably, they deliberately target water uptake $\lambda < 50$ wt%, whereas Nafion requires $\lambda \gtrsim 50$ wt% near RH ≈ 100% to reach high $\sigma$ — an explicit attempt to beat Nafion on an operational weakness rather than merely match it. Additional thresholds approximate Nafion's measured values: Young's modulus $E > 156$ MPa, $T_g > 396$ K, $T_d > 553$ K, O$_2$ permeability $< 18$ Barrer, H$_2$ permeability $\lambda < 50$0 Barrer, and band gap $\lambda < 50$1 eV (justified by the ~1.23 V water-splitting voltage plus ~2.0 eV electrode potential). Chemical constraints exclude halogens and amide groups (which trap water via hydrogen bonding, hydrolyze, and serve as radical degradation sites) while requiring sulfonate –SO$\lambda < 50$2, a restriction imposed largely because ~97.8% of curated conductivity data involves sulfonate chemistries — a data-driven constraint that limits the chemical diversity of the search.

Machine-learning models

Eight property models were trained on curated datasets ranging from 603 entries (H$\lambda < 50$3 permeability) to 8,962 ($\lambda < 50$4), using Gaussian Process Regression on hierarchical chemical fingerprints computed from SMILES representations; $\lambda < 50$5/$\lambda < 50$6 and gas permeabilities were handled with multi-task models to exploit inter-property correlations. Reported accuracies are strong:

Property	Data size	$\lambda < 50$7	Error
$\lambda < 50$8	2,462	0.84	0.12 orders of magnitude
$\lambda < 50$9	2,120	0.96	0.10 o
$\lambda \gtrsim 50$0	915	0.82	0.15 o
$\lambda \gtrsim 50$1	8,962	0.99	10.0 K RMSE
$\lambda \gtrsim 50$2	6,585	0.96	21.9 K RMSE
$\lambda \gtrsim 50$3	1,021	0.95	0.07 o
$\lambda \gtrsim 50$4	603	0.97	0.06 o
$\lambda \gtrsim 50$5	3,879	0.97	0.25 eV RMSE

For scattered properties spanning 4–6 orders of magnitude, order-of-magnitude error (OME) is used instead of RMSE; OME values of ~0.1 correspond to roughly 3–5% of the training range. The two weakest models, $\lambda \gtrsim 50$6 ($\lambda \gtrsim 50$7) and $\lambda \gtrsim 50$8 ($\lambda \gtrsim 50$9), are precisely the properties where prediction confidence matters most for candidate ranking, and the authors acknowledge this indirectly when noting relatively high uncertainty in their $\sigma$0/$\sigma$1 predictions for known membranes.

Design space construction

The design space combines ~30,000 literature-reported polymers with ~66 million synthesizable polymers generated by RxnChainer, a VFS implementation that propagates over 7 million commercially available monomers (sourced from TSCA, ZINC-22, ChemBL, eMolecules) through hundreds of rule-based polymerization templates (step growth, chain-growth addition, ring-opening, metathesis). Because VFS mimics real reaction pathways, generated polymers carry an inherent synthesizability prior. For tractability, subset 2 was restricted to four families — polyimides, polyesters, polyureas, and polyurethanes — so the "essentially unlimited" VFS space was sampled only within these chemistries, a scope limitation worth noting when interpreting the candidate set.

Validation against known membranes

Screening subset 1 rediscovered Nafion, Pemion, sPPO, and sulfonated aromatic poly(ether sulfone) copolymers (SPAES), providing an internal consistency check. Predicted $\sigma$2 curves reproduce the magnitude and temperature trend of measured data for Nafion and Pemion from the literature, and for sPPO the authors performed their own electrochemical impedance spectroscopy measurements (24–90 °C, in-plane conductivity via $\sigma$3) on commercial InnoSep-C membranes, confirming the predictions. Quantitative agreement across other properties is good: predicted $\sigma$4 K versus measured 463 K and $\sigma$5 K versus 682 K for sPPO; Nafion predictions fall within its reported ranges for all thresholded properties. Two new SPAES variants meeting all criteria were identified from a patent family by exhaustive substitution screening. The main caveat is the elevated uncertainty bands on the $\sigma$6 and $\sigma$7 predictions, which the authors attribute to insufficient training data volume for these models.

New candidates and structural analysis

Applying the full criteria stack to the 66-million-polymer space yielded 1,738 candidates: 41 from subset 1 and 1,697 from subset 2 (136 polyesters, 1,569 polyimides). Structural analysis of the candidate set reveals a consistent chemical motif: sulfonate (required), phenyl and 1,4-phenylene groups in every candidate, biphenyl in 86%, and imide groups in 89%. The authors rationalize this convergence mechanistically — rigid aromatic backbones raise $\sigma$8 via suppressed segmental mobility, strong C($\sigma$9)–C($E > 156$0) bonds (~110 kcal/mol dissociation energy) raise $E > 156$1, quasi-planar packing suppresses gas crossover, and imide dipoles improve cohesion. This coherence between the ML-selected candidates and physical structure–property reasoning lends credibility to the screen, though it also implies the candidate set is chemically homogeneous, concentrated in sulfonated aromatic polyimides.

Limitations and open questions

Several limitations temper the results. First, experimental validation covers only known membranes; none of the 1,697 novel candidates has yet been synthesized and tested, so the workflow's predictive power on genuinely out-of-distribution VFS-generated structures remains unverified — the central open question of the paper. Second, the restriction to sulfonate-containing polymers, while justified by data availability, may exclude viable non-sulfonated acid chemistries (e.g., phosphonic acids). Third, the confinement of subset 2 to four polymer families leaves most of the VFS-generated space unexplored. Fourth, durability under electrolyzer conditions — radical attack, hydration cycling, tens of thousands of operating hours — is addressed only through proxies ($E > 156$2, $E > 156$3, bond strengths), not through lifetime models. Finally, the proposed LLM-based interactive agent remains a conceptual outlook rather than a demonstrated system.

Conclusion

This work demonstrates an end-to-end, application-driven AI pipeline for PEM design: quantitative criteria derived from electrolyzer operation, multi-task Gaussian Process models with quantified uncertainty, a 66-million-polymer synthesizable design space from virtual forward synthesis, and screening validated against measured data for Nafion, Pemion, and sPPO. The identification of 1,738 halogen-free candidates — including two new SPAES compositions — is a concrete deliverable, but the strategy's ultimate value hinges on ongoing synthesis-and-testing campaigns for the novel candidates, whose outcomes will determine whether the VFS–ML workflow generalizes beyond the training distribution.