Atomistic MD Assessment of Al2SiO5 Aluminosilicate for Desalination Water/Ion Selectivity and Complementary Machine Learning Forecasting of Reverse Osmosis Performance
摘要
Despite growing global water stress, the rational design of desalination membranes remains constrained by a gap between atomistic transport mechanisms and operational performance prediction. Here, we present, to our knowledge, the first dedicated atomistic desalination transport assessment of Al2SiO5 aluminosilicate as an ultrathin reverse osmosis model membrane candidate, together with complementary machine learning regression models trained on conventional RO operational data. The MD component provides molecular-scale insight into Al2SiO5 water/ion selectivity, whereas the ML component demonstrates system-level forecasting of flux and salt rejection from routine RO monitoring variables. Because the operational dataset was not obtained from Al2SiO5 membranes, the two components are interpreted as complementary rather than directly coupled. Non-equilibrium molecular dynamics simulations were performed for 5 ns at ~ 300 K using an intentionally elevated transmembrane pressure (~ 150 MPa) to access statistically meaningful permeation events on nanosecond timescales. To ensure physically rigorous treatment of long-range interactions, the simulation cell was expanded to a 3 × 3 periodic array in the membrane plane (approximately 37.4 × 42.1 × 160.0 Å), enabling the use of a standard 12.0 Å non-bonded cutoff that satisfies the minimum image convention and is consistent with TIP4P/2005 water and ClayFF aluminosilicate force fields. The results are, therefore, interpreted mechanistically rather than as direct industrial flux predictions. The Al2SiO5 framework remained structurally intact throughout production (RMSD = 0.76 ± 0.16 Å; membrane thickness = 6.66 ± 0.16 Å), with preserved aluminosilicate coordination confirmed by persistent Si–O (~ 1.48 Å) and Al–O (~ 1.70 Å) RDF signatures. Continuous water permeation was observed across the 3 × 3 membrane, with an early-stage permeation rate of 469.5 ± 1.4 water molecules ns−1 from the 0.2 to 2.0 ns linear regime, corresponding to an MD-apparent permeance of 2.14 × 103 LMH bar−1 under accelerated driving conditions. The later reduction in slope was consistent with finite feed depletion in the nanoscale simulation cell. Ion transport was strongly suppressed, with a final ion rejection of approximately 94.2%: Out of 180 Na⁺ and 180 Cl− ions in the feed, only 9 Na⁺ and approximately 12 Cl− ions reached the permeate side over the full 5 ns production window. Apparent free-energy profiles provide a mechanistic explanation for this selectivity, revealing substantially higher barriers for Na⁺ (3.06 kcal mol−1, ~ 5.1 kBT) and Cl− (2.58 kcal mol⁻1, ~ 4.3 kBT) than for water (0.70 kcal mol⁻1, ~ 1.2 kBT) near the membrane region. At the operational scale, gradient boosting models trained on routine RO monitoring variables, including feed pressure, temperature, flow rate, and membrane age, predicted permeate flux (R2 ≈ 0.79) and salt rejection (R2 ≈ 0.85), with feature-importance analysis identifying membrane aging and hydraulic driving conditions as dominant performance determinants. Together, these findings indicate that Al2SiO5 provides a water-accessible yet ion-unfavorable model interface consistent with high intrinsic selectivity, while ML models trained on conventional RO operational data demonstrate system-level forecasting of flux and salt rejection. Direct MD-informed ML prediction of Al2SiO5 membrane performance remains a future direction requiring Al2SiO5-specific or multi-material performance datasets.