Machine learning-assisted ultrafiltration for sustainable sub-20 nm nanoparticle removal in chip production

The presence of particles in processing liquids, such as ultrapure water (UPW), isopropyl alcohol (IPA), and sulfuric acid (H2SO4) for wafer cleaning, up to 800 times in advanced semiconductor chip production, can cause defects and yield loss. Size-exclusion nanofiltration (NF) membranes with extremely small pores of 1–10 nm are widely used to remove sub-20 nm NPs during chip production, ensuring the cleanliness of these liquids, as commercial liquid particle detectors cannot quantitatively measure these tiny NPs. However, NF membrane’s small pores lead to high energy consumption. To achieve more sustainable chip production, the research team reported that ultrafiltration (UF) membranes with 20–100 nm pores can effectively capture sub-20 nm nanoparticles (NPs) by adsorption when favorable filtration conditions are met. However, due to the complex mechanisms underlying UF against sub-20 nm NPs, sustainable UF systems can be identified only with the aid of machine learning (ML). Therefore, a homemade electrospray aerosolization and particle classification system was developed to generate additional retention data for UF against 3–20 nm NPs in water and in >96 % concentrated H2SO4. A ML model was developed using high-quality data and theoretical retentions derived from the xDLVO theory to identify optimal filtration conditions for achieving sustainable UF. Results showed that 3 and 5 nm NPs could be retained at 99.9 % efficiency by a ∼50 nm-rated UF membrane in UPW, proving energy-efficient, high-NP retention by UF is feasible. In 96 % H2SO4, however, low experimental retention of ∼5 % for 20 nm SiO2 NPs by both 70 nm and 100 nm rated PTFE membranes was observed. The current successful ML model for UPW will be extended to H2SO4 when more retention data is available for smaller NPs (e.g., 5 and 10 nm) in H2SO4.