“`html

Producing superior batteries, accelerated electronics, and more efficient pharmaceuticals relies on the identification of novel materials and the confirmation of their quality. Artificial intelligence is aiding in the initial aspect, utilizing tools that sift through material catalogs to swiftly identify promising candidates.

However, once a material is manufactured, confirming its quality still necessitates examination with specialized devices to verify its efficacy — an expensive and lengthy process that can impede the advancement and distribution of new technologies.

Now, a novel AI instrument developed by MIT engineers may assist in alleviating the quality-control constraints, providing a quicker and more economical solution for specific materials-driven sectors.

In a study published today in the journal Matter, the researchers introduce “SpectroGen,” a generative AI instrument that enhances scanning functions by acting as a virtual spectrometer. The tool receives “spectra,” or measurements of a material in one scanning mode, such as infrared, and produces what that material’s spectra would appear as if examined under an entirely different mode, such as X-ray. The AI-generated spectral outputs correspond, with 99 percent precision, to the results acquired from physically scanning the material with the new device.

Certain spectroscopic modalities unveil unique properties in a material: Infrared reveals a material’s molecular structures, while X-ray diffraction displays the material’s crystalline arrangements, and Raman scattering highlights a material’s molecular vibrations. Each of these characteristics is crucial in assessing a material’s quality and usually involves complex workflows with multiple costly and separate devices for measurement.

With SpectroGen, the researchers foresee that a variety of measurements can be executed using a single and more affordable physical device. For example, a production line could perform quality assurance of materials by scanning them with a single infrared camera. Those infrared spectra could subsequently be inputted into SpectroGen to automatically produce the material’s X-ray spectra, without the factory needing to maintain and manage a separate, often pricier X-ray-scanning laboratory.

The new AI solution generates spectra in under a minute, a thousand times quicker compared to traditional methods that can take hours to days for measurement and confirmation.

“We believe that it’s not necessary to conduct physical measurements in all the modalities required, but perhaps just in a single, simple, and economical modality,” states study lead Loza Tadesse, assistant professor of mechanical engineering at MIT. “Then you can use SpectroGen to generate the remaining data. This could enhance productivity, effectiveness, and the quality of manufacturing.”

The study was spearheaded by Tadesse, with former MIT postdoc Yanmin Zhu acting as the primary author.

Beyond connections

Tadesse’s multidisciplinary group at MIT initiates technologies that enhance human and planetary health, creating innovations for uses ranging from swift disease diagnostics to sustainable agriculture.

Zhu pointed out the growing utilization of generative AI tools for discovering new materials and pharmaceutical candidates, pondering whether AI could also be employed to generate spectral data. In essence, could AI function as a virtual spectrometer?

A spectroscope investigates a material’s attributes by directing light of a specific wavelength into the material. That light causes molecular bonds within the material to oscillate in ways that scatter the light back out to the scope, where the light is captured as a wave pattern, or spectra, that can then be interpreted as a signature of the material’s structure.

For AI to produce spectral data, the conventional method would entail training an algorithm to identify relationships between physical atoms and features within a material, along with the spectra they generate. Given the complexity of molecular configurations within just a single material, Tadesse states that such an approach can rapidly become unmanageable.

“Accomplishing this even for just one material is unfeasible,” she remarks. “So, we contemplated, is there an alternative way to interpret spectra?”

The team discovered a solution in mathematics. They recognized that a spectral pattern, which is a sequence of waveforms, may be mathematically represented. For example, a spectrum that consists of a series of bell curves is known as a “Gaussian” distribution, which correlates with a specific mathematical expression, unlike a series of narrower waves, referred to as a “Lorentzian” distribution, that is characterized by a distinct algorithm. Interestingly, for most materials, infrared spectra characteristically contain more Lorentzian waveforms, while Raman spectra tend to be more Gaussian, and X-ray spectra exhibit a combination of both.

Tadesse and Zhu integrated this mathematical interpretation of spectral data into an algorithm, which they implemented into a generative AI model.

“It’s a physics-aware generative AI that comprehends what spectra are,” Tadesse affirms. “And the essential innovation is that we interpreted spectra not as a result of chemicals and bonds, but as mathematical constructs — curves and graphs, which an AI tool can grasp and interpret.”

Data assistant

The team showcased their SpectroGen AI tool on a vast, publicly accessible dataset of over 6,000 mineral samples. Each sample includes details on the mineral’s characteristics, such as its elemental composition and crystal structure. Many samples in the dataset also present spectral data in various modalities, including X-ray, Raman, and infrared. The team fed several hundred of these samples to SpectroGen, a process that trained the AI tool, also referred to as a neural network, to learn connections between a mineral’s different spectral modalities. This training enabled SpectroGen to accept spectra of a material in one mode, such as infrared, and generate what a spectra in an entirely different mode, such as X-ray, should resemble.

After training the AI tool, the researchers supplied SpectroGen with spectra from a mineral in the dataset that was not part of the training samples. They instructed the tool to generate a spectra in a different mode based on this “new” spectra. The AI-generated spectra, they found, closely matched the mineral’s actual spectra, which had been originally recorded by a physical device. The researchers performed similar experiments with several other minerals and discovered that the AI tool swiftly produced spectra, with 99 percent correlation.

“We can input spectral data into the network and obtain another completely different kind of spectral data, with very high accuracy, in under a minute,” Zhu notes.

The team asserts that SpectroGen can generate spectra for any type of mineral. In a manufacturing context, for example, mineral-based materials employed in semiconductor and battery technologies could first be quickly scanned by an infrared laser. The spectra resulting from this infrared examination could then be input into SpectroGen, which would subsequently generate a spectra in X-ray, enabling operators or a multiagent AI platform to assess the material’s quality.

“I envision it as having an agent or co-pilot, assisting researchers, technicians, pipelines, and industry,” Tadesse expresses. “We plan to tailor this for different industrial requirements.”

The team is exploring methods to adapt the AI instrument for disease diagnostics and for agricultural monitoring through an upcoming project supported by Google. Tadesse is also advancing the technology into the field via a new startup and envisions making SpectroGen accessible to a broad range of sectors, from pharmaceuticals to semiconductors to defense.

“`