Andrew Williams reports on the projects that are bringing extra intelligence to spectroscopy
There is a small but growing number of organisations around the world that are trialling the use of artificial intelligence (AI)-based (or assisted) spectroscopy techniques. So, what are the most recent developments? What AI software and spectroscopy technologies are being used? What are the current and potential applications? And what innovations and trends can we expect in this expanding sector in the coming years?
Instantaneous predictions
One of the most interesting recent initiatives is the Artificial Intelligence for Spectroscopy (ARTIST) project at Aalto University in Finland, which employs AI to speed up the spectroscopic analysis of materials and the discovery of new molecules and materials. As project lead Patrick Rinke, head of the Computational Electronic Structure Theory (CEST) group at Aalto University, explains, the idea for ARTIST was born out of the observation that conventional spectroscopy – of the experimental as well as the theoretical type – is ‘slow, expensive and often tedious.’
“Experiments require large-scale facilities or expensive equipment, and dedicated staff that maintain the equipment and have expertise in using it. Theoretical spectroscopy, that is a computational solution of the Schrödinger equation, also requires large-scale infrastructure in the form of super computers, complex software that solves the equations and expert knowledge on using the software,” he says.
“This implies that materials are often studied very thoroughly one at a time, but not on a large scale of thousands or hundreds of thousands at a time,” he adds.
However, in recognition of the fact that spectroscopy is one of the essential measurement techniques in the natural sciences, and one that has been used for decades, Rinke explains that a lot of spectroscopic data is ‘already out there in the world’ – and, in its different forms, can be used to ‘establish a relation between the structure and composition of a material and its properties.’
“This relation is encoded in the spectrum. With ARTIST we wanted to tap into this resource and train an AI to infer structure-spectra and spectra-property relationships from the world’s spectroscopy data. Properly trained, the AI could then make spectra or property predictions instantaneously without the need for arduous experiments or calculations,” he says.
Although the project is still in its infancy, Rinke reveals that the team has already built a prototype to demonstrate that it is possible for an AI to infer the spectrum, in this case a photoemission spectrum, from the structure of a material alone – in this example, relating solely to the chemical elements and ‘xyz’ positions of each atom in a molecule.
Neural network
In technical terms, Dr Milica Todorović, research fellow at the Computational Electronic Structure Theory (CEST) group at Aalto University, explains that ARTIST is a Python-based neural network code that has been trained on data relating to 100,000 small organic molecules, and which is capable of making spectra predictions of these types of molecules with 97% accuracy.
“We have used it to quickly screen a new data set of 100,000 molecules. To get the spectra for these molecules took a couple of seconds via imaging measuring in a lab. Then we screened the output for interesting molecules with interesting spectral properties,” she says.
However, according to Todorović, the application domain is ‘not limited to small molecules,’ and can be ‘trained on any substance or materials type, provided there is enough training data.’ Although not limited to such areas, the team also envisages that ARTIST could be particularly useful in application areas that require the spectral properties of materials, such as optoelectronics, plasmonics or photochemistry, for example, photocatalysis.
“ARTIST is purely software-based and uses deep neural network architectures. In our initial work, we tested three different deep neural network topologies and found that convolutional neural networks, used in image processing, and deep tensor neural networks, used in language processing, work best,” says Todorović.
During set-up, ARTIST is trained on spectroscopy data relating to pairs of molecules and the corresponding spectrum. In the training, the ‘weights’ of the neural network, as well as its hyper-parameters are determined. Once it is trained, ARTIST then receives a molecule or a material as input, in this case the xyz coordinates of the molecular geometry, and makes a prediction of the spectrum (intensity as a function of energy) instantly.
“Training ARTIST is the time consuming and difficult part that requires expert knowledge. GPU hardware is optimal for training ARTIST. Once it is trained, a laptop or a web server can make the predictions,” adds Rinke.
Earlier cancer detection
Elsewhere, UK-based medical technology outfit Lancor Scientific has created an innovative AI-assisted spectroscopic device, with which it aims to achieve 90% accuracy in cervical cancer screening. As professor Paul Darbyshire, chief technology officer (CTO) at Lancor Scientific, explains, the company’s method is ‘unique’ by virtue of the manner in which its Tumour Trace OMIS (Opto-Magnetic Imaging Spectroscopy) device is used in tandem with sophisticated AI algorithms.
“The spectroscopy is driven by quantifiable quantum physics, that detects cancer in the first instance. Only having first established that there are cancerous cells, do the AI algorithms come in to play to classify the degree of progression of the cancer. This is significantly different from most ‘AI for cancer detection’ approaches, which use AI as the ‘first sweep’ for cancer,” he says.
“Those approaches run the risk of over-fitting and over-diagnosis, leading to over-treatments and worse patient outcomes. Our approach offers the best option for reducing the burden of cancer, by finding it early,” he adds.
Deep learning tool
Ultimately, Rinke observes that AI technology ‘nicely complements’ conventional spectroscopy because it ‘fills a gap’ between experimental and theoretical spectroscopy.
“The objective of theoretical spectroscopy is often to predict the properties of materials or chemicals. But this is a slow process that AI can speed up considerably. Experiments often have the opposite objective, to learn about a material or a substance by measuring spectra.
“The challenge is to interpret the spectra and to extract information from them. This is an inverse problem, inverting the map between structure and spectrum. Here AI can help again, inferring relations, features, patterns and properties from measured data. The development of new technologies and new materials requires both fast predictions and intelligent inference,” he adds.
However, in moving towards this goal, Todorović stresses that, although humans are very good at analysing individual systems very deeply and thoroughly, they ‘cannot cope well with massive amounts of data’ – whereas AIs, in particular deep learning types, ‘thrive’ on large amounts of quality data, which ‘are not always available and would need to be generated.’
Another challenge relates to what Todorović describes as the ‘inverse problem.’ For example, she points out that inverting the structure-spectra relationship is ‘not trivial, because the inversion might generate nonsensical structures or because the relationship contains non-invertible elements.’
“Both challenges are currently addressed in data science and machine learning communities,” she says.
Moving forward, Todorović predicts that the key developments in this area in the coming years will be in the fields of materials design and interpretation of measured spectra – and she highlights the fact that ‘almost any materials scientist, be it in academia or R&D, frequently asks themselves the question what would be the best material to make their application or product better or what would give them entirely new functionality.’
“The often ‘vague’ design criteria or specifications are currently hard to translate into concrete AI algorithms, but this will improve rapidly. Also the inverse problem will become easier to solve and then AI diagnostics could assist experimental spectroscopy,” she says.
“For ARTIST, we envisage a universal spectra predictor, a web-based deep learning tool that learns continuously and dynamically and evolves somewhat autonomously. Users can enter molecules or materials and will receive an instant spectrum prediction. ARTIST will eventually support different spectroscopies and encompass many materials classes,” she adds.
