This explainer video highlights how spectroscopy is being integrated with artificial intelligence to improve detection accuracy of microplastics.
This video was created based on our recent news coverage of three studies that examined detecting microplastics in the environment. It was created using NotebookLM.
Currently, there is much research investigating the presence of microplastics in various environments. These studies have emphasized that their pervasive presence can lead to ecological harm and result in health risks to humans and other living organisms (1–3). To classify and quantify microplastics, researchers have been proposing the use of spectroscopy techniques, including Raman spectroscopy and Fourier transform infrared (FT-IR) spectroscopy.
Microplastics are everywhere but finding them with any accuracy has been a huge headache for scientists. So how do you even begin to track a pollutant that's hiding in plain sight? The problem is that they're so small; you can't see them, and that makes counting them, figuring out what kind they are, and understanding their true impact incredibly difficult.
Traditionally, researchers have relied on two main scientific tools. The first one, attenuated total reflectance Fourier transform infrared (ATR FT-IR) spectroscopy, identifies the plastic by figuring out its core chemical ingredients (2). The second one, Raman spectroscopy, can identify the plastic by its unique molecular shape. However, recent studies demonstrated that each method on its own could only correctly identify the type of plastic approximately 75% of the time (1–3). That means a full 25% was either being missed or misidentified for reliable science, which is a huge margin of error.
Therefore, researchers sought a smarter strategy, and as a result, they are incorporating AI into their work. This strategy is called data fusion, and it works on three different levels. At the low level, you're just dumping all the raw data together from the start (2). At the mid-level you pull out the most important features from each technique and then combine those data (2). And at the high level, you actually let each method make its own independent conclusion first, and then the AI steps in to make one final superior judgment based on both of them (2).
Combining the raw data, or low-level fusion, increased the classification accuracy to 88%. Using feature-level fusion improved performance further, reaching 97%. The high-level fusion approach achieved the highest accuracy, at 99%.
This represents a substantial improvement compared to earlier methods, where roughly one in four particles might have been misclassified.
Importantly, the approach was not limited to controlled laboratory samples. When applied to more complex matrices, including milk, cola, and tap water, the high-level fusion model maintained an accuracy above 98%, indicating robustness in realistic conditions.
