The production of better batteries, faster electronics and more effective pharmaceuticals depends on the discovery of new materials and verification of their quality. Artificial intelligence helps with the former, offering tools that search material catalogs to quickly flag promising candidates.
However, once the material is produced, verifying its quality still involves scanning it with specialized instruments to check its properties, a costly and time-consuming step that can hold up the development and distribution of new technologies.
Now, a new artificial intelligence tool developed by MIT engineers could help remove the quality control bottleneck by offering a faster and cheaper option for some materials-based industries.
In study that comes out today in the diary Materialresearchers present “SpectroGen”, a generative artificial intelligence tool that enhances scanning capabilities by acting as a virtual spectrometer. The tool takes “spectra,” or measurements of a material in one scanning modality, such as infrared, and generates what the spectrum of that material would look like if it were scanned in a completely different modality, such as X-rays. The spectral results generated by artificial intelligence match with 99% accuracy those obtained by physically scanning the material with the new instrument.
Certain spectroscopic modalities reveal specific properties of a material: infrared reveals the material's molecular groups, X-ray diffraction visualizes the material's crystal structures, and Raman scattering illuminates the material's molecular vibrations. Each of these properties is necessary to assess the quality of the material and usually requires painstaking work using many expensive and different measuring instruments.
Scientists predict that thanks to SpectroGen it will be possible to perform various measurements using one and cheaper physical telescope. For example, a production line could perform quality control on materials by scanning them with a single infrared camera. These infrared spectra can then be entered into SpectroGen to automatically generate X-ray spectra of the material, without the need for the factory to own and operate a separate, often more expensive, X-ray scanning laboratory.
The new AI tool generates spectra in less than one minute, a thousand times faster than traditional approaches, which can take anywhere from several hours to several days to measure and validate.
“We believe that physical measurements do not need to be made by all the necessary methods, but perhaps just one simple and cheap method,” says study co-author Loza Tadesse, an assistant professor of mechanical engineering at MIT. “Then you can use SpectroGen to generate the rest. This can improve productivity, efficiency and production quality.”
The lead author of the study is former MIT postdoctoral fellow Yanmin Zhu.
Outside of bonds
Tadesse's interdisciplinary group at MIT is pioneering technologies that improve the health of people and the planet, developing innovations for applications ranging from rapid disease diagnosis to sustainable agriculture.
“Disease diagnosis and material analysis in general typically involve scanning samples and collecting spectra in different modalities, using different instruments that are bulky and expensive and cannot be found in a single laboratory,” Tadesse says. “So we looked at how to miniaturize all this equipment and streamline the experimental pipeline.”
Zhu noted the growing use of generative AI tools to discover new materials and drug candidates, and wondered whether AI could also be used to generate spectral data. In other words, can AI act as a virtual spectrometer?
A spectroscope examines the properties of a material by sending light of a specific wavelength into the material. This light causes the molecular bonds in the material to vibrate in a way that scatters them back into the range, where the light is recorded as a pattern of waves or spectra that can then be read as a signature of the material's structure.
For AI to generate spectral data, a conventional approach would require training an algorithm to recognize connections between physical atoms and material features and the spectra they produce. Tadesse says that given the complexity of molecular structures within just one material, this approach could quickly become unfeasible.
“It's impossible to do this even with just one material,” he says. “So, we thought, is there another way to interpret the spectra?”
The team found the answer using mathematics. They realized that a spectral pattern, which is a sequence of waveforms, could be represented mathematically. For example, a spectrum containing a series of bell curves is called a “Gaussian” distribution, which is associated with a certain mathematical expression, compared to a series of narrower waves, known as a “Lorentzian” distribution and described by a separate, distinct algorithm. And as it turns out, for most materials, infrared spectra characteristically contain more Lorentz waveforms, Raman spectra are more Gaussian, and X-ray spectra are a mixture of the two.
Tadesse and Zhu turned this mathematical interpretation of the spectral data into an algorithm, which they then incorporated into a generative artificial intelligence model.
“It's a physics-literate generative AI that understands what spectra are,” says Tadesse. “And the key new thing is that we've interpreted spectra not in terms of how they arise from chemicals and bonds, but it's actually math – curves and graphs that the AI tool can understand and interpret.”
Co-pilot date
The team demonstrated the SpectroGen AI tool on a large, publicly available dataset of over 6,000 mineral samples. Each sample contains information about the mineral's properties, such as its elemental composition and crystal structure. Many of the samples in the dataset also contain spectral data in various modalities such as X-ray, Raman, and infrared. Of these samples, the team sent several hundred to SpectroGen in a process that trained an artificial intelligence tool, also known as a neural network, to learn the correlations between the mineral's different spectral modalities. This training allowed SpectroGen to take spectra of a material in one modality, such as infrared, and generate what the spectra should look like in a completely different modality, such as X-ray.
After training the AI tool, the researchers fed SpectroGen spectra of a mineral from the dataset that was not included in the training process. They asked the tool to generate spectra in a different modality based on these “new” spectra. They found that the spectrum generated by the artificial intelligence was very close to the actual spectrum of the mineral, which was originally recorded by the physics instrument. The researchers performed similar tests with a number of other minerals and found that the AI tool quickly generated spectra with a 99 percent correlation.
“We can feed spectral data into the network and get a completely different type of spectral data with very high accuracy in less than a minute,” Zhu says.
The team says SpectroGen can generate spectra of any type of mineral. For example, in a manufacturing environment, mineral-based materials used for semiconductor and battery technology can first be quickly scanned with an infrared laser. Spectra from the infrared scan can be fed into SpectroGen, which will then generate X-ray spectra that operators or a multi-agent AI platform can inspect to assess material quality.
“I think of it as an agent or co-pilot supporting researchers, technicians, pipelines and industry,” Tadesse says. “We plan to adapt it to the needs of different industries.”
The team is exploring ways to adapt the AI tool for disease diagnosis and agricultural monitoring as part of an upcoming Google-funded project. Tadesse is also putting the technology into practice through a new startup and plans to make SpectroGen available to a wide range of sectors, from pharmaceuticals to semiconductors to defense.
