Source- wwdmag.com
Smarter water, wastewater treatment & management through machine learning
Machine Learning Basics
Techniques used in the machine-learning field accelerate data analysis and pattern recognition, allowing development of predictive models that can be used to forecast the design and behaviours of parts and processes.
Unlike classic approaches for modelling, machine-learning models are not explicitly programmed by humans and are capable of evolving when introduced to new training data. These capabilities eliminate some of the guesswork that goes into problem-solving when using models and can hasten the discovery of useful insights.
Data is fundamental to machine learning. When put to work, machine-learning algorithms generate models of a system, process or part using training data—historic data compiled from theory-based calculations performed through simulations or from observation. Researchers employ trained machine-learning models to make predictions based on new, unidentified or unlabeled data.
“Machine learning is powerful because it allows us to answer questions about big data that were previously intractable or difficult to solve, using the vast amounts of data that we have now in combination with huge amounts of computing resources,” said Elise Jennings, computer scientist for the Argonne Leadership Computing Facility, a DOE Office of Science User Facility.
Optimizing Process Design
Machine-learning techniques can aid in designing wastewater treatment processes by accelerating process simulation—a critical step in developing new plants, as well as improving the design of existing plants. Today, this step is both time- and computationally intensive due to the range of parameters involved affecting incoming waste streams. Because streams can contain various contaminants and toxicants, there are numerous biochemical, mechanical and environmental factors that influence the process, and thus have to be accounted for computationally.
“Today we try to simulate the process by varying a range of design and operational parameters. This method of working involves a lot of iterative steps, involving many simulations that take a long time to do,” said environmental analyst May Wu of Argonne’s Energy Systems division. “Using machine learning is a smart way to reduce the number of high-resolution simulations that are needed, because—with just a fraction of the simulations normally done—machine-learning algorithms can adapt themselves to come up with an optimal solution, which is your design.”
Optimizing New Technologies
Wu, who works with the industry to evaluate new technologies, says that machine learning can similarly be applied to analyze emerging technologies for new or existing facilities, which could help engineers deploy these technologies much faster.
“Speed is probably the biggest value that machine learning can offer. Using these techniques, we can learn what will work, and what probably won’t work, faster than we would on our own over the same period of time,” Wu said.
Argonne demonstrated this capability through a collaboration with Aramco, a global energy company, and Convergent Science, a software company. With these collaborators, Argonne researchers used machine learning and simulation data to optimize the design of an engine to run on a new fuel.
Their work relied on a machine-learning method known as deep learning, which uses neural networks—a class of algorithms that mimic the networking of the human brain to solve problems.
With high-fidelity simulation alone, development time for the energy company added up to months. However, using neural networks and simulation data to train them, Argonne helped reduce development time to days.
Optimizing Plant Operation & Controls
After deciding how to design a plant, the next step was figuring out how to best operate it. A treatment plant that operates more efficiently can mitigate waste of important resources such as energy, which often is the single largest operating expense for water and wastewater treatment facilities.
One of the keys to improving energy efficiency is reliably predicting water demand, which is influenced by weather, consumer behaviors, and other interrelated factors. Machine-learning models trained with sensory and weather data can help predict when demand will be high or low. These insights can be used to automate machine controls, including schedules for pumping and timing for the addition or removal of chemicals or microorganisms. More efficient operation of the plant, in turn, minimizes unnecessary energy use.
Machine learning also can be used to predict process upsets, which are disruptions or failures in a normal process (e.g. a leaky pipe or equipment malfunction). Like water demand, process upsets can be caused by any number of interrelated factors, such as extreme weather or aging infrastructure. They also can have costly ramifications, at worst threatening the safety of communities and the surrounding habitat.
“Process upsets are a huge headache for us as engineers because it takes a lot of time and effort to identify the source of the problem and come up with a solution. All the while your water composition, or the organisms in the water, may be affected,” said Meltem Urgun-Demirtas, engineer and group leader for bioprocesses and reactive separations in Argonne’s Applied Materials division.
Machine learning can be used to develop models that accurately can predict when and under what conditions disruptions will occur. Facility operators can use these insights to stop upsets before they start or develop strategies to respond to them more quickly.
“If we can solve these problems faster—in moments as opposed to days or months—we can minimize the number and duration of disruptions and increase savings,” Urgun-Demirtas said.
Material Discovery & Synthesis
Another area where machine learning can have a potential impact is in the development of materials and coatings for membranes and filtration. Better materials could solve problems that create process inefficiencies, such as membrane fouling, in which particles interact or adhere to membrane surfaces or pores, hindering their performance.
However, identifying new materials and predicting their effectiveness is complex, involving many compounds—water molecules, chemical compounds, biological organisms, pollutants and more—as well as many reaction pathways.
Machine learning can help to accelerate material discovery by generating models that can predict the properties of new materials with unprecedented speed. At Argonne’s Center for Nanoscale Materials, researchers repeatedly have used machine-learning techniques in this way and have successfully characterized a number of new materials with promising energy applications.
Argonne Capabilities for Machine Learning
As a multidisciplinary research laboratory, Argonne National Laboratory is home to experts in not just machine learning, but also manufacturing, process engineering, water engineering and more. In addition, the laboratory is equipped with facilities to enable machine-learning studies across disciplines.
Among the laboratory’s premier resources for machine learning, modeling and simulation is the Argonne Leadership Computing Facility (ALCF), a DOE Office of Science User Facility that houses supercomputers and computing experts. Here, collaborators can find support for developing and optimizing machine-learning approaches and running high-fidelity simulations.
Researchers and industry around the world are actively engaging ALCF experts and resources to study everything from turbulent flows to light-absorbing materials for solar-powered windows to entire transportation networks.
“Thus far, Argonne has made tremendous strides in applying machine learning to energy- and transportation-related problems,” said John Harvey, an Argonne business development executive in Argonne’s Technology Commercialization and Partnerships division. “As we continue to build capacity in those areas, we are also looking for new opportunities, and new applications to advance American manufacturing using these tools.”