The revolutionary potential and obstacles of artificial intelligence and machine learning took center stage at the October 9 MIT Materials Day Symposium. Experts presented groundbreaking innovations in zeolite compound formation, accelerated pharmaceutical development, cutting-edge optical technologies, and numerous other advancements.
“Machine learning is transforming every aspect of materials research,” stated Carl V. Thompson, Director of the Materials Research Laboratory.
“We're increasingly collaborating with intelligent systems to guide our decisions on which materials to develop,” explained Elsa A. Olivetti, the Atlantic Richfield Associate Professor of Energy Studies. She noted that machine learning is not only helping determine how to create these materials through novel synthesis insights but is, in some cases with robotic systems, actually fabricating them.
Keynote presenter Brian Storey, director of accelerated materials design and discovery at Toyota Research Institute, discussed machine learning applications in accelerating the transition from internal combustion engines to electric vehicles. Professor Ju Li, the Battelle Energy Alliance Professor of Nuclear Science and Engineering and professor of materials science and engineering, presented innovations in atomic engineering through elastic strain and radiation manipulation of atoms.
Revolutionizing Porous Materials
Olivetti and Rafael Gomez-Bombarelli, the Toyota Assistant Professor in Materials Processing, collaborated to leverage machine learning for deeper insights into porous materials known as zeolites. These silicon and aluminum oxide compounds serve diverse applications, ranging from pet litter to petroleum refinement.
“Essentially, the concept is that the pore possesses the ideal dimensions to accommodate organic molecules,” Gomez-Bombarelli explained. While engineers recognize approximately 250 zeolites in this category, physicists can calculate hundreds of thousands of potential structural configurations. “Some can transform into one another,” he noted. “You could extract one zeolite, apply pressure or heat, and it converts into a different structure that might offer enhanced value for specific applications.”
Traditional approaches interpreted these crystalline structures as combinations of building blocks. However, when analyzing zeolite transformations, more than half the time revealed no shared building blocks between the original and transformed structures. “Building block theory offers some interesting insights but fails to fully explain the transformation rules from A to B,” Gomez-Bombarelli observed.
Innovative Graph-Based Methodology
Gomez-Bombarelli's novel graph-based approach discovered that when representing each zeolite framework structure as a graph, these graphs correspond before and after transformation in zeolite pairs. “Certain transformation classes only occur between zeolites sharing identical graphs,” he stated.
This research evolved from Olivetti's analysis of 2.5 million materials science journal articles to uncover synthesis methods for various inorganic materials. The zeolite study examined 70,000 papers. “One significant challenge in learning from literature is that we primarily publish successful outcomes and positive results,” Olivetti mentioned. In the zeolite research community, scientists also document unsuccessful attempts. “This provides us with valuable data for learning,” she emphasized. “We've leveraged this dataset to predict potential synthesis pathways for creating specific types of zeolites.”
In previous research with University of Massachusetts colleagues, Olivetti developed a system identifying common scientific terminology and techniques across this extensive literature collection, connecting similar findings. “One crucial challenge in natural language processing involves extracting these interconnected relationships throughout a document,” Olivetti explained. “We're developing tools capable of establishing these connections,” she added.
AI-Enhanced Chemical Synthesis
Klavs F. Jensen, the Warren K. Lewis Professor of Chemical Engineering and Professor of Materials Science and Engineering, described a chemical synthesis system integrating artificial intelligence-guided processing steps with a robotically operated modular reaction platform.
For those unfamiliar with synthesis, Jensen clarified that “You begin with reactants, add necessary reagents, catalysts, and other components to initiate the reaction, progress through intermediates, and ultimately obtain your final product.”
The artificial intelligence system analyzed 12.5 million reactions, establishing a rule set or library from approximately 160,000 of the most frequently used synthesis recipes, Jensen explained. This machine learning approach recommends processing conditions including catalysts, solvents, and reagents for the reaction.
“The system can utilize information from published literature regarding conditions and other parameters to formulate a recipe,” he stated. However, due to insufficient data, a chemistry expert must still specify concentrations, flow rates, process stack configurations, and ensure safety before transmitting the recipe to the robotic system.
The researchers demonstrated this system by predicting synthesis plans for 15 pharmaceutical or drug-like molecules—including the painkiller lidocaine and several antihypertensive medications—and then producing them using the system. The flow reactor system differs from batch systems. “To accelerate reactions, we typically employ significantly more aggressive conditions than batch processes—elevated temperatures and higher pressures,” Jensen noted.
The modular system comprises a processing tower with interchangeable reaction modules and various reagents, interconnected by the robot for each synthesis. These findings were published in Science.
Former PhD students Connor W. Coley and Dale A. Thomas developed the computer-aided synthesis planner and flow reactor system, respectively, while former postdoc Justin A. M. Lummiss conducted the chemistry work alongside numerous MIT Undergraduate Research Opportunity Program students, PhD candidates, and postdocs. Jensen also acknowledged contributions from MIT faculty members Regina Barzilay, William H. Green, A. John Hart, Tommi Jaakkola, and Tim Jamison. MIT has filed a patent for robotic fluid connection handling. The software suite suggesting and prioritizing potential synthesis routes is open-source, with an online version available at the ASKCOS website.
Advancing Machine Learning Robustness
Deep learning systems demonstrate remarkable performance on benchmark tasks including image and natural language processing applications, according to Professor Asu Ozdaglar, who heads MIT's Department of Electrical Engineering and Computer Science. Nevertheless, researchers remain far from fully understanding why these deep learning systems function, when they will succeed, and how they generalize. When errors occur, they can be dramatically incorrect.
Ozdaglar provided an example of an image with a state-of-the-art classifier that could correctly identify a picture of a cute pig. However, “When you introduce minimal perturbation, the identical classifier mistakenly identifies it as an airliner,” Ozdaglar explained. “This serves as an example where people jokingly say machine learning is so powerful it can make pigs fly,” she remarked to audience laughter. “This immediately demonstrates that we must advance beyond our standard approaches.”
A potential solution exists in an optimization formulation known as a Minimax or MinMax problem. Another context where MinMax formulation appears is in generative adversarial network or GAN training. Using examples of real and fake car images, Ozdaglar explained, “We want these fabricated images to follow the same distribution as the training set, achieved through two competing neural networks—a generator network and a discriminator network. The generator creates these fake images from random noise, which the discriminator network attempts to distinguish as real or fake.”
“This represents another MinMax problem where the generator minimizes the distance between fake and real distributions, while the discriminator maximizes that distance,” she noted. The MinMax problem approach has become fundamental to robust training of deep learning systems.
Ozdaglar added that EECS faculty are applying machine learning to new domains, including healthcare, citing Regina Barzilay's work in breast cancer detection and David Sontag's research utilizing electronic medical records for medical diagnosis and treatment.
The EECS undergraduate machine learning course (6.036) enrolled 800 students last spring and consistently maintains 600 or more enrollees, making it MIT's most popular course. The new Stephen A. Schwarzman College of Computing offers an opportunity to create a more dynamic and adaptable structure than MIT's traditional departmental framework. For instance, one proposal involves creating several cross-departmental teaching groups. “We envision offerings such as computing foundations, computational science and engineering, social studies of computing, with these courses attended by all students and jointly taught by faculty across MIT,” she explained.
Optical Innovations
Juejun "JJ" Hu, associate professor of materials science and engineering, detailed his research integrating a silicon chip-based spectrometer for detecting infrared light wavelengths with a newly developed machine learning algorithm. Traditional spectrometers, dating back to Isaac Newton's first prism, operate by splitting light, which reduces intensity. However, Hu's version captures all light at a single detector, preserving intensity but creating the challenge of distinguishing different wavelengths from a single capture.
“To resolve the trade-off between spectral resolution and signal-to-noise ratio, you must employ a new spectroscopy tool called wavelength multiplexing spectrometer,” Hu explained. His novel spectrometer architecture, known as digital Fourier transform spectroscopy, incorporates tunable optical switches on a silicon chip. The device functions by measuring light intensity at different optical switch settings and comparing results. “Essentially, you have a system of linear equations providing various linear combinations of light intensity at different wavelengths as detector readings,” he stated.
A prototype device with six switches supports 64 unique optical states, providing 64 independent measurements. “The advantage of this new device architecture is that performance doubles with each additional switch,” he noted. Collaborating with Brando Miranda at MIT's Center for Brains Minds and Machines, he developed a new algorithm, Elastic D1, achieving resolution down to 0.2 nanometers and delivering accurate light measurements with only two consecutive readings.
“We believe this unique combination of new spectrometer hardware and algorithm can enable numerous applications from industrial process monitoring to medical imaging,” Hu stated. Hu is also applying machine learning in his research on complex optical media such as metasurfaces—novel optical devices featuring arrays of specially designed optical antennas that introduce phase delays to incoming light.
Poster Session Recognition
Nineteen MIT postdocs and graduate students delivered two-minute presentations about their research during a poster session preview. At the Materials Day Poster Session following the symposium, award recipients included mechanical engineering graduate student Erin Looney, media arts and sciences graduate student Bianca Datta, and materials science and engineering postdoc Michael Chon.
The Materials Research Laboratory supports interdisciplinary faculty, staff, and student groups, funded by industry, foundations, and government agencies, to conduct fundamental engineering research on materials. Research areas encompass energy conversion and storage, quantum materials, spintronics, photonics, metals, integrated microsystems, materials sustainability, solid-state ionics, complex oxide electronic properties, biogels, and functional fibers.
