Learn more about Purency and microplastics

The analysis based on machine learning algorithms offers several key advantages over conventional methods of data analysis: Database-based analysis solutions often have a couple of hundred IR spectra in their library. The cross-checking of each pixel to the library is therefore quite slow and gets even slower as the library gets bigger. A more robust classification requires a bigger library which will slow down the process of the classification even further. With machine learning, however, you do not have a trade-off between quality and speed of the analysis: the Microplastics Finder makes the analysis more robust and the results more precise. All while being faster. A lot faster.

Our software runs on ordinary office PCs. However, please ensure that you have at least 16 GB of RAM. We recommend 32 or 64 GB of RAM since hyperspectral images can become very large (e.g. 20 GB). As our software is enhanced for multi-screen use, we highly recommend that you have two screens connected to your workstation.

The Microplastics Finder is optimized for FT-IR imaging. However, we are always broadening our horizon. We are currently running tests for processing Raman data as well and enlarge the Microplastics Finder's capabilities. If you are using a Raman instrument and are interested in using our software, please contact us to explore the opportunities.

The Microplastics Finder is a software product that automates the data analysis of microplastics measurements using FTIR imaging. Thus, an IR (imaging) microscope is needed for its usage. The real power of the Microplastics Finder lies within its ability of processing FTIR images fast and deliver robust classification results.

Microplastics Finder covers 99% of microplastics by covering the 20 most abundantly produced polymers. If you are looking to identify other particles, please contact us and we will explore the options. Generally speaking, the model would be trained to cover another type of particle and then a new model would be deployed.

We have a built-in feature that states how certain the prediction algorithm is about its classification of the particle. With high values, you can be sure that the particle was classified correctly. With low values, you are able to manually check whether the classification was really correct by using the built-in spectral reference database.

The Microplastics Finder classifies pixels. The size of particles that can still be detected, therefore, depends on the spatial resolution of the setup used.

Currently, we offer the classification of 20 most abundantly produced polymer types. This covers more than 99% of current microplastics. As the Microplastics Finder purely focusses on finding microplastics, every non-polymer pixel that is identified will be treated as matrix.

The results provided by the Microplastics Finder are reproducible and transparent. Meanwhile the method eliminates the need for time-consuming manual rework, thereby, minimising subjective bias and maximising the comprehensibility.

Our algorithm was trained by experts with over 12 000 spectra of microplastic particles in a vast number of matrices and on several different filter types. By design, this methodology ensures that the algorithm is extremely robust and classifies even noisy environmental samples reliably. A performance that a conventional database with a couple of hundred spectra cannot achieve.

The decision-making process of the algorithm to classify and assign each pixel to a predefined polymer class is an extremely fast process. It is an order of magnitude faster than other particle software. For example, using database-based approach means each pixel's spectrum is compared to every spectrum of the database, thus, the process takes longer the more spectra the database contains. Therefore, there is a trade-off between time and analytical quality which you do not have with a machine learning approach.

Spectroscopic approaches, and imaging techniques in particular, are very promising when it comes to analysing a sample for microplastics. However, the enormous amounts of complex data pose a challenge. A powerful tool was needed to reliably identify microplastics, even in complex samples: Based on years of experience we crafted the Microplastics Finder with the goal to raise the analysis of microplastics to the level of routine analysis. Microplastics analysis becomes scalable and new types of questions about the presence and origin of microplastics can be answered. Therefore, enabling risk assessments and quality control in an efficient manner.

An FT-IR image is densely packed with a ton of molecular information and the experimental procedure for data collection is quite easy. However, the enormous and complex amounts of data that are produced, make it difficult to obtain the relevant information about quantity, size and type of the polymer particles in the sample. Here, the Purency Microplastics Finder comes into play as it analyses the complete data set within a few minutes. This guarantees that no information gets lost and the full potential of FT-IR imaging is unlocked.

The Microplastics Finder checks each pixel of the hyperspectral image and assigns it to a predefined polymer class. This decision-making process is done by our built-in machine learning algorithm, which uses the characteristic regions of the spectra to distinguish the polymer types. After the entire image has been processed, neighbouring pixels with the same polymer class are labelled as particles. Further, properties such as the polymer type and the size of each particle are visualized in a clearly arranged results table.

At Purency, we use machine learning to automate certain processes that are too time-intensive to be done by humans. Machine learning models are ultimately mathematical models which were created from training data by a machine learning algorithm. So, to solve the microplastics classification problem, we trained a model to decide which pixel is a polymer and which one is not. We also taught it to distinguish between different types of polymers. There are different approaches to train machine learning models: supervised and unsupervised learning. In unsupervised learning you simply feed data into the algorithm and the algorithm learns by itself what the outcome should be. In supervised learning, the algorithm is trained with labelled data, which is the outcome of the classification process. Purency uses supervised learning and trains the model based on representative data which was compiled and analysed by microplastics experts around the world. Thereby, providing a fully trained, out of the box, and robust method to laboratories analysing microplastics – enabling you to identify not just what one person, but a group of microplastics experts would find in a sample.