The integration of machine learning and artificial intelligence has become increasingly prevalent in materials science investigations. At MIT, materials science and engineering associate professor Juejun "JJ" Hu pioneered an algorithm that significantly boosts the functionality of chip-based spectrometers, while Atlantic Richfield Professor of Energy Studies Elsa A. Olivetti engineered an artificial-intelligence framework that meticulously analyzes scientific publications to uncover materials science "formulas."
These distinguished MIT faculty members, along with featured presenter Brian Storey—who serves as the director of accelerated materials design and discovery at Toyota Research Institute—will share valuable perspectives and revolutionary developments in their machine learning-enhanced research during the MIT Materials Research Laboratory's yearly Materials Day Symposium scheduled for Wednesday, Oct. 9 in Kresge Auditorium.
Associate Professor Hu recently shared the inspiration behind his groundbreaking spectrometer innovation and expressed confidence about machine learning and artificial intelligence becoming standard instruments in materials exploration.
Q: Your spectrometer research specifically incorporated machine learning methodologies. How is this innovative approach transforming the discovery process in materials science?
A: Essentially, we've pioneered a novel spectrometer technology that enables the miniaturization of large components onto a compact silicon chip while preserving exceptional performance. We've engineered an algorithm that facilitates information extraction with a significantly enhanced signal-to-noise ratio. We've successfully validated this algorithm across various spectrum types. By comparing two repeated measurements, the algorithm distinguishes different light wavelengths while minimizing the influence of measurement noise. The algorithm achieves a 100% improvement in resolution beyond the conventional Rayleigh limits typically found in textbooks.
Q: How are you leveraging machine learning to identify innovative optical materials and configurations for your mid-infrared lens work utilizing optical antenna arrays?
A: We're partnering with researchers at UMass [the University of Massachusetts] to create a deep learning algorithm for designing "metasurfaces"—a category of optical devices that utilize arrays of specially engineered optical antennas to introduce phase delay to incoming light, rather than relying on traditional geometric curvature to construct elements like lenses. This approach enables us to accomplish diverse functionalities. A significant challenge with metasurfaces has been that designers historically relied on trial-and-error methods.
We've established a deep learning algorithm that learns from existing datasets. Through this training process, the algorithm gradually becomes "intelligent." It can evaluate the viability of irregular shapes that extend beyond conventional forms like circles and rectangles. The algorithm can identify subtle relationships between complex geometries and electromagnetic responses—connections that are typically non-obvious—and it can uncover these hidden patterns more rapidly than traditional comprehensive simulations. Additionally, the algorithm can eliminate material and function combinations that are destined to fail. With conventional approaches, researchers would waste extensive time exploring the entire design space before reaching this conclusion, but our algorithm provides this information almost instantaneously.
Q: What other developments are promoting the adoption of machine learning in materials science?
A: Another significant trend we're observing is the dramatically improved accessibility to commercially available, cloud-based computational facilities with tremendous processing power. This convergence of hardware, accessibility, powerful computing resources, and innovative algorithms is what empowers us to achieve new breakthroughs. Returning to the metasurface example, if you examine earlier designs, researchers primarily employed regular geometric shapes like circles, squares, and rectangles. However, our team, along with many others in the field, is now transitioning toward topologically optimized optical devices. The combination of advanced algorithms and formidable computational capabilities is essential for designing extensive structures such as macroscopic, three-dimensional, topologically optimized optical systems.
