Abstract
<jats:p>ChemGNN, a deep-learning-accelerated framework that learns a graph-based surrogate of membrane transport and actively proposes high-performance designs. Each CNT configuration is encoded as a heterogeneous graph in which tube nodes carry geometry and functionalization features, edges encode inter-tube spacing, and global features describe operating conditions such as pressure, temperature gradient, and salinity. A contextual message-passing encoder produces node embeddings that a transformer-style readout fuses into a representation predicting water permeance and salt rejection. An expected-improvement acquisition function then selects the most informative candidates for molecular-dynamics validation, closing an active-design loop that concentrates expensive simulations on promising regions of the design space. We evaluate ChemGNN on a dataset of 12,480 configurations with 86 experimental hold-out measurements, comparing against an MD-only random search, a Gaussian-process surrogate, and a topologyagnostic multilayer perceptron surrogate. ChemGNN achieves the highest water permeance of 63.5 LMH/bar and salt rejection of 99.2 percent while requiring roughly an order of magnitude fewer molecular-dynamics evaluations than all baselines. Extensive analyses further reveal the non-linear sensitivity of performance to tube diameter, array pitch, temperature gradient, and feed salinity, and a human evaluation by domain experts confirms the novelty and practical relevance of the proposed designs. These results demonstrate that graph-aware deep learning effectively captures the coupled geometry-transport relationship underlying CNT membrane desalination.</jats:p>