Nanolayers Research Computing Milestone of Machine Learning
The EU-funded project MAST3RBoost is attempting to create hydrogen (H2) storage tanks so H2 can be used as a fuel for road vehicles. For this, MAST3RBoost is creating innovative materials that can withstand high pressures and low temperatures, needed to hold enough H2 in a small volume to be used as vehicle fuel.
The truth is, we make that sound simple, but it is not. One of the challenges of finding good materials is that there are many critical parameters that are hard to evaluate and even visualize. These materials are nanoporous, with microscopic pores of 100 nanometres or less, so they can hold H2. They are created by using a starting mix of materials and making them react physically and chemically. Afterwards, their properties must be characterised, and their performance evaluated to determine if it can be used as is, improved, or discarded, as a material for the H2 tank. But how can researchers understand the relationships between all the parameters in 100 different materials and how the starting building blocks and conditions determine the final properties of the material created?
Nanolayers Research Computing Milestone of Machine Learning
MAST3RBoost partner Nanolayers Research Computing had a key milestone in their progress in the project, as they have managed to use machine learning to solve some of these challenges.
Using data to predict the properties of materials from starting inputs
The first half of Nanolayers’ breakthrough was made by training a regression model to predict the characteristics of a new material using the known synthesis parameters and starting components.
Through a statistical analysis method called principal component analysis (PCA) and machine learning, Nanolayer created and trained a model to do just that. The model was initially trained with only 80 samples and did a good job of predicting some of the parameters with least variability, but less so for the parameters that varied highly from sample to sample. For those, the team had to train the model with new materials produced later. After that, even high variability parameters could be predicted accurately.
Visualizing material properties and their similarities
The other half of the breakthrough used machine learning to help researchers visualize all the complex data from the materials already tested and their characteristics.
For this, a machine learning algorithm called t-distributed stochastic neighbour embedding (t-SNE) was used, which projects high-dimensional, hard to visualize data points onto a 2D or 3D space while preserving their similarity relationships. The resulting 2D or 3D maps show underlying patterns between samples, allowing to understand data in a better way, so researchers can make the best use of their results.
Data processing makes complex data understandable
The breakthroughs mentioned above might seem trivial, but visualizing and understanding complex data can help tremendously. Long are the days when scientists only sat at a bench doing straightforward experiments. The amount of data that can be generated nowadays is hard to comprehend even for experts in their fields.
The ability of machines to compute these data and apply statistical analysis to correlate parameters is more and more needed to deal with the data overload. By using these machine learning techniques, the MAST3RBoost applies state-of-the-art computational techniques to understand the likely outcomes of experiments before they are done, and to correlate and understand the properties of materials created.
Visualizing material properties and their similarities
While the MAST3RBoost project has generated a lot of data by itself, much more is needed to train machine learning algorithms better. The work of the project establishing these models and algorithms is the first step of many, within and without the project.
Machine learning is rapidly changing the way researchers approach complex challenges in materials science. As more data becomes available and machine learning models are refined, the potential to design, optimise, and discover next-generation materials only increases, and this allows the next breakthroughs that might change society, like a small fuel H2 tank for vehicles.
Author: Darío Sánchez