CuspAI is betting that the next leap in environmental remediation won’t come from a single breakthrough chemical, but from a new way of discovering them. In an interview, the UK-based science start-up’s co-founder, Max Welling, described how the company is using artificial intelligence to design molecules intended to tackle one of the most stubborn pollution problems of our time: “forever chemicals” in water—commonly referring to PFAS, a large family of per- and polyfluoroalkyl substances known for their persistence in the environment and resistance to breakdown.
PFAS are difficult not only because they don’t degrade easily, but because they behave in ways that make them hard to capture with conventional treatment. Many PFAS are highly stable, can bind to surfaces, and often travel through water systems long after they enter them. That combination—chemical stability plus mobility—has turned PFAS into a global challenge for regulators, utilities, and communities. Yet the same properties that make PFAS so problematic also create a clear target for materials science: if you can engineer molecules or materials that interact strongly with PFAS, you can potentially trap them, concentrate them, and remove them more effectively.
What CuspAI is trying to do is accelerate the search for those interactions. Instead of relying on slow cycles of synthesis and testing, the company uses computational approaches to explore chemical space and generate candidate molecules designed with specific properties. The goal is to find materials that can better target PFAS in real-world water conditions—where the chemistry is messy, competing ions and organic matter are present, and performance must be consistent rather than theoretical.
The unique angle in CuspAI’s approach is not simply “using AI for chemistry,” but using AI as a discovery engine that can propose candidates faster than traditional trial-and-error. In practice, that means building models that learn patterns from data—both experimental results and computational predictions—and then using those models to generate new molecular structures likely to meet desired criteria. The company’s work sits at the intersection of machine learning, molecular modeling, and physical chemistry, where the challenge is to ensure that the outputs aren’t just plausible on paper, but meaningful for real interactions with contaminants.
To understand why this matters, it helps to consider what “designing molecules for PFAS removal” actually entails. PFAS removal isn’t one single mechanism. Depending on the material and process, PFAS can be adsorbed onto surfaces, captured within pores, bound through specific chemical interactions, or separated through processes that exploit differences in affinity. Each mechanism has its own requirements. A molecule that binds strongly in a lab solvent might fail in water. A candidate that captures PFAS under ideal pH might lose performance when water chemistry shifts. And even if a material captures PFAS, it must do so without becoming a secondary hazard or creating byproducts that are equally difficult to manage.
That’s where AI-driven design becomes more than speed. It can help incorporate multiple constraints simultaneously—properties like binding affinity, selectivity, stability, and manufacturability—so that candidates are filtered early rather than after expensive synthesis. In other words, the model doesn’t just guess a structure; it tries to optimize toward a set of goals that reflect the realities of deployment.
Max Welling’s description of CuspAI’s work emphasizes the acceleration of candidate generation. Traditional discovery pipelines can take years: propose a hypothesis, synthesize a compound, test it, analyze results, refine the hypothesis, and repeat. Even with high-throughput experimentation, the number of possible molecules is enormous. Chemical space is vast enough that brute-force search is impractical. AI offers a way to navigate that space intelligently—learning which regions are more likely to contain useful solutions and focusing computational effort there.
But there’s a second, equally important part of the story: verification. AI can propose candidates, yet chemistry is governed by physics and thermodynamics. If the model is wrong, the candidate fails. So companies like CuspAI must connect the AI layer to scientific evaluation. That typically means using simulations and property predictors to estimate how a candidate might behave, followed by experimental testing to confirm performance. The best systems are iterative: experimental results feed back into the model, improving future proposals. Over time, the system becomes more accurate not because it “knows” PFAS in a human sense, but because it learns from the outcomes of previous attempts.
In the context of PFAS, the interaction problem is particularly challenging. PFAS molecules often have fluorinated chains that are both hydrophobic and chemically inert. Their behavior in water is shaped by the balance between the fluorinated portion and any functional groups at the head. Materials designed to capture PFAS must therefore contend with strong tendencies toward phase separation, surface adsorption, and complex interactions with other dissolved substances. Selectivity is also crucial. A material that captures PFAS indiscriminately might foul quickly or remove beneficial compounds, while a material that is too selective might miss certain PFAS variants that appear in different industrial contexts.
CuspAI’s focus on designing molecules with targeted properties suggests an attempt to address these trade-offs systematically. Rather than searching for a single “magic” compound, the company’s strategy aligns with a broader trend in clean-tech: build platforms that can adapt. PFAS contamination is not uniform. Different sites have different PFAS profiles depending on industrial history, manufacturing inputs, and disposal pathways. A platform that can generate new candidates for different PFAS mixtures could be more valuable than a one-size-fits-all solution.
This is where the “molecules” framing matters. Many PFAS remediation efforts focus on macroscopic materials—filters, resins, membranes, and sorbents. But those materials are often built from underlying chemical motifs. If you can design the molecular components that drive adsorption or binding, you can tune the performance of the larger material. AI can therefore function as a bridge between molecular design and engineering outcomes: propose molecular structures, predict their behavior, and then translate promising candidates into materials that can be manufactured and integrated into treatment systems.
There is also a strategic reason to pursue AI-enabled molecular design now. Regulatory pressure around PFAS has intensified across multiple countries, and utilities are under increasing scrutiny to demonstrate effective removal. At the same time, public awareness has grown, and communities want solutions that are not only effective but scalable. That creates a demand for remediation technologies that can be deployed widely, not just proven in pilot studies. AI-driven discovery can shorten the timeline from concept to candidate, potentially reducing the lag between scientific understanding and practical implementation.
Yet the path from candidate molecule to real-world remediation is not straightforward. Even if a molecule shows strong affinity for PFAS, it must be stable under operational conditions. Water treatment environments can involve varying temperatures, pH levels, and competing chemical species. Materials must resist degradation and maintain performance over time. They also must be regenerable or disposable in a way that does not create new environmental burdens. For PFAS, the stakes are high: capturing the contaminants is only half the job; the captured PFAS must be handled safely afterward.
This is why the “forever” aspect of PFAS is so central to the conversation. PFAS are persistent, but they are also diverse. Some PFAS are more prevalent than others, and their chemical structures vary. A remediation technology that works for one PFAS may not work equally well for another. That diversity makes the design problem multi-dimensional. AI can help by exploring a wider range of candidates and by optimizing for selectivity across different PFAS types, rather than assuming a single target structure.
CuspAI’s approach, as described by Welling, reflects a broader shift in how science start-ups are tackling complex environmental problems. Instead of treating AI as a separate tool, they treat it as part of the research loop. The model proposes, the lab tests, the results inform the model, and the cycle continues. This “closed-loop” mindset is particularly powerful in domains where experiments are expensive and time-consuming. PFAS remediation research fits that profile: testing materials against PFAS can require specialized setups, careful handling, and rigorous measurement to avoid misleading results.
Another dimension worth noting is the difference between predicting properties and designing molecules. Predicting is about estimating what already exists. Designing is about generating what doesn’t yet exist. Those are related but not identical tasks. Generative models and optimization frameworks can propose new structures, but they must be constrained by chemical validity. A model that generates impossible molecules is useless. A model that generates valid molecules but ignores synthesis feasibility is also limited. The best systems incorporate constraints that reflect both chemistry and practicality—ensuring that candidates are not only theoretically interesting but also potentially producible.
In the PFAS context, there’s also the question of how to evaluate “removal” in a way that translates to real treatment. Laboratory tests might measure adsorption capacity under controlled conditions, but real water contains a mixture of contaminants. Performance metrics can include how much PFAS is removed per unit mass of material, how quickly it is removed, how the material behaves over repeated cycles, and whether the material releases PFAS back into the water under certain conditions. AI can help optimize toward these metrics, but it requires data that reflects them. That means the quality of the experimental pipeline becomes part of the AI pipeline.
CuspAI’s emphasis on accelerating the creation of potential materials suggests that the company is working to reduce the time between hypothesis and candidate. In a field where the cost of failure is high, speed can be a competitive advantage—but only if speed doesn’t compromise scientific rigor. The most credible AI-enabled chemistry efforts are those that treat computational outputs as hypotheses to be tested, not as final answers.
There’s also a narrative shift happening in environmental tech: from reactive cleanup to proactive prevention and targeted remediation. PFAS contamination is often the result of historical use and disposal practices. Even as regulations tighten, legacy contamination remains. That means remediation will be needed for years, possibly decades. AI-enabled discovery can help by producing better materials faster, but it also supports a longer-term vision: designing molecules and materials that can address emerging pollutants before they become widespread.
In that sense, CuspAI’s