> ## Content Index
> Fetch the complete content index at: https://research-pop.com/llms.txt
> Use this file to discover other available public pages before exploring further.

# Georgia Tech researchers put machine-learning membrane choices through an experimental gate
- URL: https://research-pop.com/georgia-tech-researchers-put-machine-learning-membrane-choices-through-an-experimental-gate/
- Published: 2026-10-01T12:11:55.000Z
- Updated: 2026-10-01T12:11:55.000Z
- Author: ResearchPOP

![](https://storage.ghost.io/c/7c/5e/7c5e0911-739b-4a35-94eb-3e810bf78ba7/content/images/2026/10/Screenshot-2026-10-01-at-14.06.29.png)

Source: [https://doi.org/10.1021/acs.est.5c15241](https://doi.org/10.1021/acs.est.5c15241?ref=research-pop.com)  
At a glance

Designing a nanofiltration membrane means navigating many interacting choices: monomer chemistry, concentrations, additives, polymerization time and curing conditions. Those variables must be coordinated while balancing water permeability against salt rejection, two properties that often move in opposite directions.

Nohyeong Jeong, Elif Demirel, Changyoon Jun and Yongsheng Chen at the Georgia Institute of Technology developed a machine-learning-assisted inverse-design workflow, abbreviated MLAID. Rather than beginning with a membrane recipe and asking what performance it might deliver, the workflow begins with a desired salt-water permeability and ranks experimentally accessible combinations that could approach it.

XGBoost models learned from membrane-fabrication data collected from the literature. SHAP analysis showed which variables contributed to the models’ predictions, and Bayesian optimization ranked monomer and processing combinations for target permeabilities from 4 to 12 LMH/bar. The researchers then fabricated and tested seven prioritized conditions in the laboratory.

Four non-PEI membranes formed active layers and delivered measured permeabilities of 3.5, 6.0, 8.2 and 12.4 LMH/bar while rejecting 95.2–99.4% of Na₂SO₄. Two formulations containing polyethyleneimine, or PEI, emphasized opposite sides of the permeability–rejection balance, while one high-permeability non-PEI condition failed to form an active layer. Together, the outcomes show how model-guided ranking and experimental feedback can work as successive stages of membrane design.

## Background

Nanofiltration occupies the separation range between ultrafiltration and reverse osmosis. Its thin selective layer allows water to pass while retaining solutes according to a combination of size, charge and chemical interactions. Applications include water softening, pollutant removal, resource recovery and the separation of dissolved ions or organic molecules.

A useful membrane must transport water quickly enough to limit the required pressure and membrane area, while maintaining sufficient rejection of the target solute. Water permeability is commonly reported in litres per square metre per hour per bar of applied pressure, abbreviated LMH/bar. Increasing permeability can reduce energy or equipment demands, but opening transport pathways too far can lower selectivity.

Many commercial and research nanofiltration membranes use a polyamide selective layer prepared by interfacial polymerization. An amine dissolved in water reacts with an acyl chloride in an organic solvent at the boundary between the two phases. Because the reaction occurs rapidly at that interface, relatively small changes in monomer identity, concentration, ratio, additives, contact time or thermal treatment can alter cross-linking, surface morphology, charge and effective pore structure.

This creates a large experimental design space. Testing every possible combination is impractical, and improvements in one metric can come at the expense of another. Machine learning offers a way to extract statistical patterns from earlier studies and use them to prioritize the next experiments. Literature data, however, are heterogeneous. Different laboratories use different supports, filtration cells, salts and reporting protocols, and relatively few published membranes occupy the highest-permeability region.

The authors use machine learning as a decision-support tool rather than a replacement for fabrication. The model narrows and orders the search space. Laboratory experiments determine whether the suggested chemistry forms a usable selective layer and whether its measured permeability and rejection justify further development.

## Research question

Can a literature-trained and interpretable machine-learning workflow begin with a desired salt-water permeability, prioritize experimentally actionable monomer and fabrication combinations, and use laboratory results to inform the next design cycle?

The study addresses this question with separate models for salt rejection and salt-water permeability, SHAP attribution, Bayesian optimization, fabrication of seven prioritized conditions, cross-flow Na₂SO₄ testing and physical characterization of the four successful non-PEI membranes.

## Inside the study

![](https://storage.ghost.io/c/7c/5e/7c5e0911-739b-4a35-94eb-3e810bf78ba7/content/images/2026/10/Screenshot-2026-10-01-at-14.06.41.png)

The workflow began with nanofiltration and reverse-osmosis fabrication data collected from papers published over the preceding 20 years. The input variables included monomer identities, fabrication conditions, ion properties and hashed Morgan fingerprints that encode molecular substructures in a machine-readable form.

Before training, the researchers removed 12 records with curing times longer than 200 minutes and 84 records lacking monomer SMILES strings. They then trained separate XGBoost models for salt rejection and salt-water permeability. Keeping the targets separate allowed the workflow to examine the familiar permeability–selectivity tension rather than collapse both outcomes into one unexplained score.

![](https://storage.ghost.io/c/7c/5e/7c5e0911-739b-4a35-94eb-3e810bf78ba7/content/images/2026/10/Screenshot-2026-10-01-at-14.06.50.png)

The salt-rejection model used 513 training points. On the held-out test set, it reached R² = 0.87, a root-mean-square error of 11.9% and a mean absolute error of 7.5%. The permeability model used 347 training points and reached R² = 0.63, a root-mean-square error of 1.8 LMH/bar and a mean absolute error of 1.2 LMH/bar. Each training set represented 80% of the available target-specific data, with stratified five-fold cross-validation used during training.

The difference between the two model performances reflects the available data as well as the difficulty of the targets. Errors increased in the high-permeability region, where published examples were sparse. Variations among laboratories also introduce influences that are not fully represented by the recorded variables. The models are best read as tools for prioritizing candidates within the assembled data domain, not as universal predictors for any membrane chemistry or test arrangement.

![](https://storage.ghost.io/c/7c/5e/7c5e0911-739b-4a35-94eb-3e810bf78ba7/content/images/2026/10/Screenshot-2026-10-01-at-14.07.01.png)

SHAP analysis was used to inspect how the trained models distributed importance among their inputs. For salt rejection, prominent attributions included anion valence, the amine-to-acyl-chloride molar ratio, cation valence and acyl-chloride concentration. For water permeability, curing time, polymerization time, monomer molar ratio and additive concentration were among the influential variables.

The opposite associations of monomer ratio in the two models reflect the trade-off recorded in the literature data. A formulation or processing choice that favours a tighter selective layer may help rejection while limiting water transport. The analysis makes those learned patterns visible, but it does not establish that changing one variable in isolation will cause the corresponding performance change.

![](https://storage.ghost.io/c/7c/5e/7c5e0911-739b-4a35-94eb-3e810bf78ba7/content/images/2026/10/Screenshot-2026-10-01-at-14.07.19.png)

The authors apply the same caution to individual descriptors. Curing temperature, for example, is correlated with time and concentration choices across published studies, so its model attribution does not reproduce a simple direct experimental trend. Hashed fingerprint bits can highlight useful structural patterns, but they are statistical features rather than uniquely identified chemical mechanisms.

Bayesian optimization then searched for candidates near target permeabilities of 4, 6, 8, 10 and 12 LMH/bar. The search began with 310 monomers from the National Institute for Materials Science database. Structural-similarity, commercial-availability and experimental-compatibility filters narrowed that set to candidates that could reasonably be purchased and tested. Trimesoyl chloride, hexane and an operating pressure of 5 bar were fixed so that other choices could be compared within a shared experimental framework.

The decisive stage was laboratory fabrication. Four non-PEI candidates formed active layers and gave permeabilities of 3.5 ± 0.4, 6.0 ± 0.1, 8.2 ± 0.4 and 12.4 ± 1.0 LMH/bar. Their corresponding Na₂SO₄ rejections were 99.4 ± 0.2%, 96.4 ± 1.0%, 98.2 ± 1.0% and 95.2 ± 2.8%.

The membrane that reached 12.4 LMH/bar came from the search aimed at a 10 LMH/bar target. It is a useful high-permeability candidate, but not an exact target match. This distinction illustrates what inverse design means in the present study: the workflow ranks promising experiments near a desired performance region rather than guaranteeing a recipe with a predetermined measured value.

The PEI-containing formulations and the failed membrane provide equally important boundary cases. One PEI membrane reached 98.3 ± 0.3% Na₂SO₄ rejection but only 1.5 ± 0.1 LMH/bar permeability. The other reached 14.7 ± 2.5 LMH/bar but retained only 59.2 ± 7.4% of Na₂SO₄. A separate non-PEI condition proposed for the 12 LMH/bar target did not form an active layer.

![](https://storage.ghost.io/c/7c/5e/7c5e0911-739b-4a35-94eb-3e810bf78ba7/content/images/2026/10/Screenshot-2026-10-01-at-14.07.27.png)

The authors relate these outcomes partly to how the input representation handles polymeric PEI and to the limited amount of training data near the highest target. A monomer-level fingerprint does not fully describe a polymer’s molecular-weight distribution, branching or behaviour during interfacial polymerization. The failed and imbalanced candidates show why experimental validation must remain inside the workflow.

Physical characterization then examined why the four successful non-PEI membranes transported water differently. As measured permeability rose from 3.5 to 12.4 LMH/bar, electron microscopy showed a change from finer, more compact surface nodules toward coarser and less densely packed domains. Cross-sectional images showed broadly comparable active-layer thicknesses, making thickness alone an insufficient explanation for the performance range.

Neutral-solute rejection measurements gave molecular-weight cutoffs of 199, 255, 275 and 357 Da in the same order as increasing permeability. Mean pore size also increased from 0.25 to 0.31 nm. These observations provide a physical basis for faster water transport: the higher-permeability membranes contained progressively more open effective pathways.

Salt rejection did not follow pore size alone. At pH 7, the four membranes had zeta potentials of −21.8, −3.2, −12.7 and −3.6 mV. The membrane with 8.2 LMH/bar permeability rejected more Na₂SO₄ than the 6.0 LMH/bar membrane despite having a larger mean pore size. Its more negative surface charge is consistent with stronger electrostatic exclusion of the divalent sulfate anion.

The characterization thus connects membrane performance to both size sieving and Donnan exclusion. Pore dimensions influence which species can enter the selective layer, while fixed surface charge changes the partitioning of ions. The measured outcome emerges from both effects rather than from one structural number.

## Takeaways and outlook

The study’s contribution is an experimentally gated design workflow rather than a single record-setting membrane. Literature data support two predictive models. SHAP makes the models’ statistical patterns inspectable. Bayesian optimization ranks candidates for target permeability regions, and laboratory fabrication determines which suggestions produce functional membranes.

Within one Na₂SO₄ benchmark and the reported test conditions, four non-PEI candidates spanned 3.5–12.4 LMH/bar while maintaining 95.2–99.4% rejection. The PEI trade-offs and the condition that failed to form an active layer are not side notes. They reveal gaps in the data and molecular representation that can guide the next iteration.

The authors propose expanding the workflow with larger and more balanced datasets, descriptors better suited to polymeric monomers, uncertainty-aware optimization, additional salts and mixtures, and performance targets such as fouling resistance and chemical or mechanical stability. The resulting design cycle is deliberately collaborative: the model proposes, the laboratory verifies, and successful as well as unsuccessful experiments improve the next search.

## About the researchers

Nohyeong Jeong and Elif Demirel (Georgia Institute of Technology) are equal first authors. Yongsheng Chen (Georgia Institute of Technology) is the corresponding author. Changyoon Jun is also affiliated with the Georgia Institute of Technology.

Jeong implemented the machine-learning model, Demirel conducted the experimental work, Jun led model validation and scientific interpretation during revision, and Chen supervised the study.

## Original research

Nohyeong Jeong; Elif Demirel; Changyoon Jun; Yongsheng Chen. “Learning the Path to Ion Separation: A Proof-of-Concept for Machine Learning-Assisted Inverse Design (MLAID) of Nanofiltration Membranes.” *Environmental Science & Technology* (2026). Published online 30 September 2026\. DOI: [10.1021/acs.est.5c15241](https://doi.org/10.1021/acs.est.5c15241?ref=research-pop.com).

---

## Research POP Notes

This article reflects the independent interpretation of the Research POP team and does not represent the views of the authors, their institutions or the journal. If you identify any inaccuracies or have concerns regarding the content, figures or attribution, please contact us at [team.researchpop@gmail.com](mailto:team.researchpop@gmail.com). We will review the matter promptly and make corrections or remove the relevant material where appropriate.