Neural Operator optimizes design of medical catheter

Cuts bacterial contamination by two orders of magnitude

Designing new devices in various engineering and scientific applications is still a trial-and-error process, driven by intuition that requires deep domain expertise and is limited by the time and cost of physical experiments. Numerical simulations present an alternative to physical experiments but are usually infeasible for complex real-world scenarios since they require an inordinate amount of computation and specialized skills to run the simulations. AI can have a transformative impact by significantly accelerating scientific simulation and modeling, as well as directly optimizing for specified design goals. We were able to demonstrate this by generating a novel optimized design of a medical device, a catheter.

A catheter is a flexible tube to draw fluids out of a human body. Unfortunately, catheters frequently cause bacterial infections by bacteria swimming upstream against the fluid flow and infecting the patient. This is, in fact, the most common healthcare-associated infection source and cause of secondary bloodstream infections. Despite all the advances in diagnosis, prevention, and treatment, it remains a severe healthcare burden, especially as antibiotic resistance rates are alarmingly high.

Instead of complicated ways to reduce infection, we went back to fundamentals. We looked at designing a geometric control of microbial distribution, a passive mechanism that is simple and cheap to construct. This involves adding asymmetric shapes, such as triangles inside the tube, that prevent bacteria from swimming upstream due to vortices and turbulence created in the fluid from irregular boundary shapes. However, finding the optimal shape of triangles that minimizes bacterial contamination to the greatest extent is not straightforward.

To achieve the goal of optimal geometric design, we studied the physical mechanism of bacteria swimming upstream to fluid flow. This is modeled through mathematical equations, known as stochastic partial differential equations (SPDE), that consider both the dynamics of fluid movement and the geometric effects of internal shapes within the catheter pipe. However, using a traditional solver to simulate the flow under different geometric shapes is slow and also does not provide any design guidance.

To overcome this challenge, we used our Neural Operator framework to significantly speed up the fluid flow simulations. But not just that, the use of Neural Operators also helps to directly optimize for the geometric shape due to differentiability. The differentiability of Neural Operators means they do not just make physical simulations faster, but also create new designs by pointing out how a design needs to be changed to better achieve its physical goals. The Neural Operator-based AI model directly improved the design and proposed an optimized design candidate with the best geometry to prevent bacteria from swimming upstream. We 3D-printed and lab-tested the design and saw up to 2 orders of magnitude reduction in bacterial contamination, reducing the overall risk of catheter-associated infections. The AI method came up with a novel optimized geometric design by predicting in-flow bacterial dynamics and optimizing to reduce bacterial contamination.

The geometric control of microbial distribution that we chose is a passive mechanism that is cheap and effective, as opposed to more complex designs that actively change the fluid flow, such as valves. It is also safer than antibiotics or other chemical methods regarding antibiotic resistance.

The paper is published in Science Advances.

Background: Neural Operators

While there have been many attempts to use large language models (LLMs) for scientific problems with some success, there are inherent limitations. LLMs are only trained on text data which provides a high-level understanding but lacks the ability to internally simulate scientific processes such as fluid dynamics or material deformation. Further, image generation models only operate at a fixed resolution, which makes them unsuitable for capturing multi-scale processes that occur across resolutions. To overcome this, we created Neural Operators that work at all resolutions, and can do super-resolution without the need for retraining. Neural Operators have replaced traditional simulations while being tens or even hundreds of thousands of times faster across multiple scientific domains, including weather forecasting, Carbon Capture and Storage modeling, industry-scale automotive aerodynamics, and early detection of disruption in nuclear fusion.