Clustering using Neural Networks

A solvent dataset is fed to the AI algorithm (a Self-Organizing Map). We can create a subset based on our 553 solvent set, or use the clients own dataset. The algorithm clusters the solvents with similar physical properties and characteristics. We perform 1000 runs on the dataset and determine which solvents are statistical significant neighbors. A shortlist is created and ranked automatically along the HS&E parameters so that you immediately see the HS&E impact of every alternative. For a more in depth explanation, read our publication in the Molecules special issue on Bio-Based and Safer Solvents.

553 solvents and 22 parameters

Our database contains 553 solvents and 22 parameters, many of those are not included in other solvent selection software or databases, such as surface tension, refractive index, density, viscosity, flashpoint, autoingition temperature, … The database only includes experimental data.

Sustainable Solvents

Our database contains the ‘Green Solvents’, but also includes common organic solvents. The list of 'green alternatives' is rather small but often it is possible to find a solvent that is healthier, or safer, without being green in all aspects.

A Neural Network to Cluster Solvents

The Artificial Intelligence module is a Neural Network. We applied the ‘Self-Organising Map’ or SOM of Kohonen, a Finnish AI researcher. Based on the physical properties only (numerical data solely) similar solvents are clustered. Link to the article here.

The SOM is based on how our visual brains perceive lines and how they categorise these lines. All the properties of the solvents are coded in a list (a vector). These vectors are offered to the network, a kind of cerebral cortex one could say. These vectors are firstly placed in a random order, but via a certain logic, similar vectors become neigbors. And this learning process comes to an end when the vector stop moving. The system then converges to an equilibrium (homeostasis) and it stops learning.

The process described above is performed 250 till 1000 times, that is what we call clusterings. The results of each clustering (the map) is stored. And then finally we perform a statistical analysis to determine which are the best clusterings. We then analyze the clusters of the 100 best performing clusterings and we look for the most frequent neighbor of each solvent.

An extra advantage of the neural network is that it learns. When new solvents are added in the database, then then knowledge model becomes smarter, it makes new connections, which is not possible with a static database.