Managing Water Supplies With Machine Learning

From wet benches to cooling systems, fabs use vast amounts of water — millions of gallons per day at a typical foundry. In this era of climate change, though, water supplies are becoming less reliable and municipal water systems are becoming more restrictive. For example, local utilities might restrict a fab’s ability to draw from the public water supply, or might supply only treated wastewater rather than potable water. Advanced water management is essential for business continuity.

Water management considers both input water supplies and output waste streams, as well as the fab’s ability to reuse its own wastewater. On the input side, fabs may need to accept less than ideal supplies, such as wastewater or ocean water. On the output side, fabs seek to recover both usable water and process chemicals. As Prakash Govindan, co-founder and COO of Gradiant Technologies explained, in both cases the first step is for the fab to understand its water supply and segregate sources that require different levels of treatment. Segregating water with high solids content, corrosive chemicals, and so on can substantially improve water management efficiency. A fab might have as many as fifteen different waste streams, each with different treatment requirements.

Segregated waste streams also facilitate chemical recovery. For example, Govindan said, isopropanol is a leading candidate for chemical recovery efforts. It’s much more expensive than water, and more abundant in fab wastes than metals like copper.

Purification technologies like reverse osmosis are more cost effective if recoverable chemicals are extracted and solids are removed first. Reverse osmosis works by passing the water to be purified through a semi-permeable membrane. Contaminants remain on the pressurized side of the membrane, while pure water diffuses through. “We can purify anything” by reverse osmosis, Govindan said, but more difficult purification tasks drive up the cost and energy requirements.

To match purification resources to the influent wastestreams, Gradiant creates a digital twin of the facility being optimized, supported with real-time sensor data from the actual facility. Machine learning models use both historical and real time information to anticipate the need for membrane replacements and other maintenance. In a microelectronics facility in Singapore, Gradiant was able to recover nearly 90% of influent wastewater, with commensurate reductions in the plant’s freshwater requirements.