Every pharmaceutical you swallow, every catalyst that speeds a chemical reaction, every solvent purified for reuse—all passed through an industrial separation process. Most likely, that process involved boiling. Heat the mixture. Collect the vapor. Repeat.

It's energy-hungry. Wasteful. And in an era fixated on emissions, increasingly indefensible.

But what if membranes could do the job instead? Thin polymer films that let solvents pass while holding back dissolved molecules. No heat required. Just pressure. The promise is enticing: lower energy, fewer emissions, costs that might actually compete. The trouble is knowing when membranes win. For which chemical? In which solvent? At what rejection rate?

Until now, that knowledge lived in scattered papers, incomplete datasets, and the intuition of a few specialists.

A new study changes that. Researchers compiled nearly 10,000 nanofiltration measurements into a single dataset. They trained a machine learning model to predict how well membranes reject specific molecules. Then they modeled the energy demands of membrane separation, evaporation, and liquid extraction across millions of chemical scenarios.

The result? A roadmap. Not for replacing all thermal separations—that's neither realistic nor necessary—but for identifying where membranes offer genuine advantages. In roughly 40% of industrially relevant cases, the authors estimate, introducing nanofiltration could cut both energy consumption and carbon dioxide emissions by similar margins.

That's not marginal. It's structural.

Assembling the pieces

Nanofiltration works by selectivity. A membrane allows solvent molecules to permeate while retaining larger or more chemically distinct solutes. The measure of this retention is rejection: a value between zero and one. High rejection means the solute stays put. Low rejection means it slips through.

Rejection depends on membrane material, solvent type, solute structure, pressure, and temperature. The design space is vast. Too vast for trial and error.

The researchers began by curating literature data. They focused on commercially available membranes and small organic molecules. The resulting dataset—called NF-10K—spans 1,089 solutes across different solvents and membranes.

They compared this with 124,000 organic compounds drawn from pharmaceutical, dye, catalyst, pollutant, and natural product databases. Chemical space visualization confirmed overlap. The training data was representative. A graph neural network learned the patterns.

The model achieved a root mean squared error of 0.12 with 89% accuracy on unseen data. Not perfect—membrane measurements themselves vary—but reliable enough for screening.

Energy in, energy saved

Knowing rejection values is one thing. Translating them into energy and emissions savings is another.

The team built mechanistic models for three separation technologies: triple-effect evaporation, continuous nanofiltration, and liquid-liquid extraction. They simulated binary separations (one solute, one solvent) and ternary separations (two solutes, one solvent). They accounted for feed concentration, target purity, and external heat integration.

For binary concentration—taking a dilute solution and making it concentrated—nanofiltration offered an average 36% energy reduction compared to evaporation alone. Even with 75% heat integration (where waste heat is recycled), nanofiltration still delivered 29% savings.

A threshold emerged. When solute rejection exceeds 0.6, nanofiltration generally beats evaporation. Below that, evaporation holds the advantage. The threshold depends on concentration and heat integration, but it provides a simple decision rule.

Solvents with high boiling points and high average rejections—water, N,N-dimethylformamide—saw the largest gains. Nanofiltration for solvent recovery worked in only 11% of cases, requiring exceptionally high impurity rejection.

Ternary separations showed even greater promise. These involve separating two solutes from each other, not just concentrating one. Here, rejection selectivity matters. The membrane must reject one solute much more than the other.

In 74% of ternary separations, nanofiltration outperformed liquid-liquid extraction. Average energy reduction: 61%. In some pharmaceutical purifications, emissions dropped by 90%.

The authors mapped these reductions globally. Countries with cleaner electricity grids saw greater emissions benefits. Membrane separations use electricity; evaporation uses heat. The carbon footprint of each differs by region. Operating cost reductions varied similarly, driven by electricity and natural gas prices.

But membrane costs matter. High upfront prices can erase savings. The researchers calculated threshold membrane prices—the maximum annual cost per square meter at which nanofiltration remains economically viable. For some applications, current prices exceed that threshold. Membranes need to get cheaper or last longer.

Searching the chemical space

With predictive models and energy calculations in hand, the researchers conducted a virtual screen. They estimated 7.1 million rejection values across industrial sectors.

They found that 59% of solutes could be concentrated efficiently using nanofiltration. That percentage rose to 86% for certain membrane-solvent combinations. About 42% of solutes showed rejection above 0.6 with solubilities of at least 10 grams per liter—enough to make concentration practical.

Solutes divided along interesting lines. Products—pharmaceuticals, catalysts, natural products—tended toward high rejection. Impurities—genotoxins, pollutants, surfactants—skewed lower. That disparity is good news for ternary separations. High rejection selectivity means better purification.

Around 65% of product-impurity pairs showed rejection selectivity high enough to at least double relative purity in a single stage. In 17% of cases, a tenfold increase was achievable. In 4%, fiftyfold.

Pharmaceuticals stood out. Roughly 91% could be removed from wastewater using nanofiltration. Genotoxic impurities, plant protection products, surfactants, and dyes also showed high retention.

Chemical substructures mattered. Large, non-polar groups correlated with high rejection. Small, polar groups with low rejection. Beta-lactam rings—common in antibiotics—consistently showed high retention. So did quinuclidine rings in catalysts.

The researchers noted a gap. Current membrane development focuses on maximizing solvent permeance and individual solute rejection. Less attention goes to rejection selectivity—the ability to distinguish between structurally similar molecules. Yet that distinction is what enables fine chemical separations.

They proposed refocusing research. Design membranes that excel at separating chemically similar solutes. Test them on diverse sets of molecules spanning both high and low rejection. Prioritize selectivity over absolute retention.

Proving it works

Three case studies validated the approach.

Nanostar sieving. This technique grows peptides on a central hub molecule, forming a "nanostar." Nanofiltration removes impurities while retaining nanostars. Previous studies reported rejections between 0.75 and 0.90—good, but not great. Low selectivity undermined efficiency.

The researchers used the machine learning model to redesign the hub. They replaced flexible ether arms with rigid amide groups. Predicted rejection jumped to 0.987 and 0.9999 for two new designs. They also swapped the coupling agent, decreasing its rejection. The combined changes increased rejection selectivity 810-fold. Experimental measurements confirmed the predictions. A "no-loss diafiltration" scenario became possible.

Apixaban synthesis. This blood thinner was the sixth top-selling pharmaceutical in 2022. The final synthesis step uses vacuum distillation to remove ammonia and methanol. The researchers simulated replacing distillation with nanofiltration. Predicted apixaban rejection: 0.963. Measured: 0.989. Energy demand dropped 95%. Carbon emissions fell 88–95%, depending on electricity source. Cost reductions were achievable in regions with favorable membrane prices and electricity costs.

Metoprolol synthesis. This beta-blocker is widely prescribed. Its final synthesis step also uses vacuum distillation and extraction. Nanofiltration predictions matched experimental results. Energy reduction: 61%. Emissions reduction: 60%. But no cost savings at current membrane prices. Membranes would need to cost less or last longer to compete economically.

These examples underscore a theme. Nanofiltration works—chemically, energetically, environmentally. Economically, it depends.

Where this leads

The authors don't claim membranes should replace all thermal separations. That's neither possible nor necessary. But they do argue for a shift in decision-making. Instead of defaulting to evaporation or extraction, engineers should evaluate membranes first. The tools now exist to do so.

The methodology combines data-driven rejection predictions with mechanistic energy models. It identifies parameter thresholds—rejection values, selectivity boundaries—that simplify technology selection. It maps energy and emissions reductions geographically. It enables virtual screening before a single experiment.

Limitations remain. Predictions falter for molecules far outside the training data. Industrial conditions are more complex than simulations assume. Life cycle assessments would provide fuller comparisons. But the framework is extensible. More data improves predictions. More detailed models refine comparisons.

The pharmaceutical industry stands to gain significantly. High-value products. Complex purifications. Stringent purity requirements. Mounting pressure to reduce waste and emissions. Nanofiltration aligns with all of it.

Water treatment is another target. Removing pharmaceuticals, pesticides, and micropollutants from wastewater. Purifying solvents for reuse. Concentrating catalysts for recovery. The study estimates that 91% of pharmaceuticals could be filtered from wastewater.

Broader adoption hinges on two factors. First, membrane costs must fall or durability must improve. Second, research must prioritize selectivity. The ability to separate structurally similar molecules matters more than maximal rejection of any single solute. Current development trends favor the latter. They should favor the former.

The study also highlights an often-overlooked connection. Sustainable energy and energy-efficient technology are linked. Nanofiltration uses electricity. Its environmental benefit depends on how that electricity is generated. Renewable grids amplify emissions reductions. Fossil-heavy grids diminish them. Decarbonizing industry requires decarbonizing energy supply.

None of this happens without data. The NF-10K dataset is open-access. So is the predictive model. Engineers can test scenarios online. Researchers can extend the work. Transparency accelerates progress.

The bigger picture

Industrial separations account for a significant fraction of global energy consumption. Improving them isn't flashy. It doesn't generate headlines. But it matters. Efficiency at scale compounds.

This study offers a methodology, not a mandate. It identifies opportunities, not guarantees. Whether industry adopts nanofiltration depends on economics, infrastructure, and inertia. The first two can be addressed. The third is harder.

But inertia isn't immovable. Data helps. So do clear thresholds, mapped reductions, and validated predictions. Engineers need tools, not rhetoric. This work provides tools.

The question isn't whether membranes can compete with evaporation. In many cases, they already do. The question is whether we'll build the systems to let them.

Credit & Disclaimer: This article is a popular science summary written to make peer-reviewed research accessible to a broad audience. All scientific facts, findings, and conclusions presented here are drawn directly and accurately from the original research paper. Readers are strongly encouraged to consult the full research article for complete data, methodologies, and scientific detail. The article can be accessed through https://doi.org/10.1038/s41560-024-01668-7