Inorganic phosphors unlocked the potential of LEDs to produce white light, revolutionising commercial lighting. Despite an enormous number of known inorganic compounds, however, finding compounds which could host rare earth elements to become new inorganic phosphors is tedious and time-consuming. Could advances in machine learning guide better material discovery? Jakoah Brgoch at the University of Houston, Texas, USA, thinks so. He and his team created a series of machine learning algorithms to identify candidate materials from a database of over 100,000 compounds.
If you look at how we spend our time online, you could feasibly make the case that robots – or artificial intelligence (AI) at least – have already taken over. Machine learning (ML)algorithms affect the movies and websites we are recommended, the products we are shown, and the opinions with which we are presented. While early advancements in machine learning were driven by the likes of Facebook, Amazon, Netflix and Google, ML is capable of improving more than just advertising revenue.
One developing field of research is the application of machine learning to aid the discovery of new materials. This approach, key in the field of ‘materials informatics’, has the potential to find ‘candidate materials’ more quickly, and can also consider materials which might otherwise have been ignored.
Making up for lost light
Professor Jakoah Brgoch and his team at the University of Houston, USA, are researching the application of machine learning techniques in the design of new lighting. Their area of interest is in luminescent inorganic compounds, materials first used in television tubes and fluorescent lamps that have now found application in computer and phone screens, and even lasers.
Perhaps their most important application is in LED lighting. Modern LED lights are 75% more efficient than filament bulbs, last up to 25 times longer, do not degrade from repeated switching, and are composed of environmentally benign materials.
The initial challenge in creating LED lighting was finding a method of creating white light. LEDs only produce one ‘band’ of colour, and white LED lights containing a mixture of red, green and blue LEDs are restrictively expensive to produce. The trick with LED lights, however, is that they do not only rely on LED technology – they actually make use of a class of materials called ‘inorganic phosphors’. These materials absorb some of the light given out by a blue or UV LED and then glow at a slightly longer wavelength. The mixture of colours from the LED and the inorganic phosphor, if paired correctly, appears white. A common combination is the use of blue LEDs with an inorganic phosphor that absorbs the blue light and emits yellow and red. However, new, useful types of these important light conversion materials are in short supply. Whilst there is a good understanding of how to tune these materials’ properties, the process of discovery itself still relies on a ‘trial and error’ approach and simple design rules.
Can algorithms do chemistry?
Professor Brgoch is interested in using machine learning to advance the field of inorganic phosphors among other functional materials. There are only a handful of usable phosphors known today that have the required properties to make them suitable for LED lighting applications. As well as the ability to absorb and emit photons at the right wavelengths, usable inorganic phosphors need to meet another crucial criterion: they need to have a high luminescent quantum yield. In other words, they need to efficiently convert absorbed photons into emitted photons.
So what properties of new materials affect their quantum yield? Perhaps the most important factor is the structural rigidity of the compound at an atomic level. Molecules with structures that are flexible tend to convert a lot of their absorbed photons into vibrations (phonons), which transfer the energy into heat rather than light.
For machine learning to use rigidity as a metric, it is important to have a single metric which can summarise the rigidity of a compound. Debye temperature is considered by many to be the best proxy currently available for rigidity, but it has a problem: it is calculated using the elastic moduli of a compound. These elastic moduli have only been measured or computationally modelled for around 10% of all known inorganic compounds.
ML-predicted Debye temperatures
This problem – the low percentage of potential materials with experimental or calculated Debye temperatures – is where Brgoch and his team used machine learning. By training an algorithm using known Debye temperatures, alongside multiple other metrics available for inorganic compounds, Brgoch’s team was able to predict the Debye temperature for a much larger number of compounds – over 120,000 – based on data from the Pearson’s crystal database.
One key factor is the availability of the electronic ‘band gap’ present in the host crystal structure. Inorganic phosphors are composed of a crystal structure doped with a very small amount of a rare earth ion. One common example of a host crystal structure is yttrium aluminium garnet, or YAG. Replacing a very small number of the ions in the crystal’s structure with a rare earth like cerium(III) produces an inorganic phosphor that can be used in white LED lighting.
It is important for the host crystal to have a large electronic band gap because small band gaps can cause lower quantum yields in the phosphor compound. So the team also created a machine learning model that could predict a material’s band gap.
Visualising the trends
Filtering the original 120,000 compounds from the database based on the machine learning-predicted Debye temperature and band gap, as well as the presence of at least three elements in the compound, left a slightly more manageable shortlist of 35,371. However, plotting the resulting predicted Debye temperatures against the band gap for each of the compounds created a sorting chart that could be used to separate ideal candidate inorganic phosphor compounds based on their position on the chart – those higher and to the right having a large machine learning-estimated Debye temperature and a large band gap in the crystal host: both good indicators of high quantum efficiency.
The resulting chart also highlights some interesting trends related to types of phosphor hosts. For example, all of the sulphides had relatively small Debye temperatures and relatively small band gaps, meaning they are less likely to be good candidate materials. On the other hand, nitrides tend to have moderately high band gaps and high Debye temperatures. It is no surprise, therefore, that many known useful phosphors are based on nitride crystal structures.
Silicates and aluminates are the most common classes of phosphor hosts, and this matches wide band gaps seen on Brgoch’s chart. These classes of phosphor hosts are also known to have a large range of quantum efficiency, and this matches the large range of predicted Debye temperatures seen on the chart. Phosphates tended to have large band gaps and moderate Debye temperatures. Recent developments have found some rare earth-substituted phosphates with good luminescent quantum efficiency, and this data supports the possibility of finding more.
The number of borates found in the very top right of the chart was something of a surprise for Brgoch. Their high Debye temperature – a proxy for crystal rigidity – was possibly because of the dense packing of atoms made possible by boron’s very small size.
Synthesis of the machine-chosen compound
To test the capability of their machine learning-driven prediction and filtration methodology, Brgoch and his team used the chart to identify a particularly remarkable compound – NaBaB9O15 – a borate with very promising predicted properties.
Brgoch and his team synthesised this compound with varying amounts of Europium, replacing between 0.5% and 5% of the barium atoms in the crystal structure. They found that these compounds could be readily prepared with a few heating steps.
Once the synthesis was complete, Brgoch’s team determined whether their predictions and assumptions were correct. Computational analysis of the structure now confirmed by X-ray diffraction showed that the Debye temperature was predicted to within 12% of its actual value, and that the host does in fact have a very wide band gap.
Europium-doped NaBaB9O15 was shown to absorb UV light, and the brightest of the tested compounds contained 3% Europium, corresponding to a fantastic luminescent quantum efficiency of 95%. Concentrations greater than 3% Europium led to a decrease in quantum efficiency – thought to be due to a self-quenching process where Eu atoms are too close to each other in the crystal structure, leading to transfer of energy between Eu atoms.
High temperatures? No sweat
NaBa0.97Eu0.03EuB9O15, unlike many inorganic phosphors, also retains its high quantum efficiency even at high temperatures. This is an important property for inorganic phosphors, which can be subjected to high temperatures, even under normal operating conditions. Further investigation reveals that the structure of this compound is not modified even up to 500 Kelvin. This is why the high quantum efficiency is not altered even at elevated temperatures.
This compound is activated by ultraviolet light, so it may currently be better-suited to laser applications than white LED lighting. However, altering the properties of this compound by doping the same crystal host with different rare earth metals in different ratios is known to be an effective method of altering these types of properties. Changing the synthetic conditions can also change where the rare earth is located in the crystal structure. Indeed, the researchers have found that reducing the reaction temperature to make this material produces a highly coveted green-emitting phosphor that is also functional at high temperature.
More important is the fundamental result that machine learning can be successfully employed to extend data sets, which has led to the discovery of a new viable inorganic phosphor material.
- Zhuo, Y., Mansouri Tehrani, A., Oliynyk, A. O., Duke, A. C., & Brgoch, J. (2018). Identifying an efficient, thermally robust inorganic phosphor host via machine learning. Nature Communications, 9(1), 1.
- Zhuo, Y., Mansouri Tehrani, A., & Brgoch, J. (2018) Predicting the band gaps of inorganic solids by machine learning. The Journal of Physical Chemistry Letters, 9(7), 1668−1673.
- Duke, A. C., Hariyani, S., & Brgoch, J. (2018). Ba3Y2B6O15:Ce3+—A High Symmetry, Narrow-Emitting Blue Phosphor for Wide-Gamut White Lighting. Chemistry of Materials, 30(8), 2668–2675.
- Zhong, J., Zhuo, Y., Hariyani, S., Zhao, W., Wen, J., & Brgoch, J. (2019). Closing the Cyan Gap Toward Full-Spectrum LED Lighting with NaMgBO3:Ce3+. Chemistry of Materials, 32(2), 882–888.
- Zhuo, Y., & Brgoch, J. (2021). Opportunities for Next-Generation Luminescent Materials through Artificial Intelligence. The Journal of Physical Chemistry Letters, 12(2), 764–772. DOI: 10.1021/acs.jpclett.0c03203
- Zhuo, Y., Hariyani, S., Armijo, E., Abolade Lawson, Z., & Brgoch, J. (2019). Evaluating Thermal Quenching Temperature in Eu3+-Substituted Oxide Phosphors via Machine Learning. ACS Applied Materials & Interfaces, 12(5), 5244–5250.
Jakoah Brgoch and his team are researching the application of machine learning techniques in the design of new lighting.
The National Science Foundation (DMR-1847701 and CER-1911311), the Welch Foundation (E-1981), the Texas Center for Superconductivity and the Alfred P. Sloan foundation.
Prof Jakoah Brgoch completed his Ph.D. at Iowa State University and postdoctoral research at the University of California, Santa Barbara. Jakoah joined the University of Houston Department of Chemistry in 2014 and is leading a multidisciplinary research group covering topics that span materials synthesis, characterisation, computation, and machine learning.
3585 Cullen Blvd. Room 112
University of Houston
Houston, TX 77204 USA