Suitcases and pipes: Making machine learning work for clean water Part II

Last time we looked at how machine learning can help water utilities manage their maintenance and operations efforts - especially when dealing with hard-to-reach parts of the water system like buried water pipes. Today, let’s talk about how machine learning is being used in developing new technologies and building prototypes for decentralized, small-scale systems.


Desalination has been studied and deployed at scale for several years now. As different parts of the planet face increasing water stress, desalination is being evaluated as one of several potential solutions - together with water conservation and recycled water. In the Middle East of course, large-scale desalination plants have been in operation for several decades, with Israel being one of the countries at the forefront of developing and implementing the technology.


Large-scale desalination plants have the advantage of scale - you can build a single system and then connect it to your existing water network. Sufficient data are available so that machine learning can be applied effectively to produce useful solutions. But what about smaller systems? Can machine learning be used in such systems and what problems can they be used to solve?


Portable decentralized desalination systems are an example of small scale systems that are now gaining wider acceptance. They are being used in areas where building a pipe infrastructure is challenging and during disasters in order to ensure access to clean drinking water. Most portable desalination systems use high-pressure pumps to push water through filters - a technique that can be challenging to miniaturize. 


Researchers at MIT, however, used novel processes to replace current filtration technology and built a portable desalination unit that costs less than $50, requires less power to operate than a cell phone charger and produces drinking water that exceeds WHO’s water quality standards. The technology and prototype were designed using machine learning and controlled using a smartphone. 


The scientists used a combination of ion concentration polarization (ICP) and electrodialysis (ED) to remove salts, bacteria, viruses and other impurities from the sea water. Instead of filtering water, ICP applies an electrical field to membranes placed above and below a channel of water. The membranes repel positively or negatively charged particles -- including salt molecules, bacteria, and viruses -- as they flow past. The charged particles are funneled into a second stream of water that is eventually discharged. 


However, ICP doesn’t remove all the salts floating in the middle of the channel - and this is where the ED process comes into play. In electrodialysis, DC power is used to move charged ions through a membrane so that clean water flows out of one end and concentrate water with the salts flows out of another end. Both ICP and ED systems require less power as they use low-pressure filters and direct current. 


The challenge in building a system with multiple processes is that an optimal configuration must be found. This is true even where only two processes are in play. In other words - how many ICP systems and ED systems are needed and how must they be connected to achieve the maximum water output that meets water quality standards. And this is where machine learning comes into play!


Before the optimal configuration is determined, we need data on the different states of the system and the outputs associated with them. The scientists ran several experiments using different configurations of the ICP and ED systems and recorded their results (salt removal ratio and energy per ion removal). They also defined functions to represent the intermediate state of the configurations (voltage and salinity of each stage) - something that is difficult to measure in practice. These data were used to train a machine learning model which then predicted the optimal configuration of the system. The model predictions were once again tested in the experimental set-up to verify their accuracy.


The researchers also tested a variety of machine learning models (support vector regression, linear regression, gaussian process regression and a feedforward neural network) to identify the ones that were most useful. Given the relatively small dataset that they were working with, linear regression - which is one of the simplest models - performed the best. In other words, if you have a relatively simple system (in this case, there are only 2 processes) and a small dataset, you’re probably best served by using simple models as opposed to complicated models like neural networks. This is a situation that comes up quite frequently in earth and environmental systems - which is why data science and machine learning need to be used appropriately - not just applied to every problem!


The other interesting aspect of the system the MIT researchers developed was their use of solar panels to power the portable system and a smartphone controller to monitor the electricity use and salinity results. That’s another part of data science in action - building and connecting devices wirelessly so that it’s easy to collect data and monitor the system in action. In the early stages of using data science in a sector, this is probably one of the most common applications.


If you'd like to learn more about data science, machine learning and their applications in the water sector - join the waitlist for our water tech sector and get notified about new courses and the latest in science and technology in the space.

Comments

Popular posts from this blog

Moonshots, Models, IoT and Machine Learning in Agriculture

Machine Learning, AI and Climate

Experiences in Smart Water - the Singapore Story