Creating the Architecture of Materials by Design

Northwestern researcher Wei Chen has helped pioneer the use of AI to identify promising materials for energy technologies.

Materials informatics: teaching optimization to handle real chemistry

Materials design research often involves choosing among distinct options, such as which element or crystal structure to use. Because these are separate categories rather than points along a smooth scale, standard optimization methods have difficulty comparing them and predicting which combinations will work best. Chen’s group developed a new approach, the latent variable Gaussian process (LVGP). This machine learning model learns from existing data to predict how different combinations of material choices and numerical factors will perform. Its distinctive feature is that it can work with categorical choices—such as which element, class of crystal structure, or additive to use—by converting them into numerical representations that the model can compare.

LVGP helped Chen and collaborators discover new metal-insulator transition compounds with potential applications in electronics, energy storage, and thermal management. The same machine learning engine has since been redeployed for dielectric nanocomposites, gas-capturing metal-organic frameworks, and solar-cell materials. The team has since enhanced LVGP to handle the scale that real materials databases demand. They created a fully Bayesian version for better-calibrated uncertainty, and a stochastic-variational variant built for many-level categorical variables and large sample sizes.

Big data for materials and chemistry: fixing bias before it compounds

As databases tied to the Materials Genome Initiative grew, and as Northwestern's Center for Hierarchical Materials Design (CHiMaD), where Chen is a principal investigator, expanded its holdings, a subtler problem emerged: size does not guarantee representativeness. Chen and her collaborators tackled this directly, using active learning to steer future data collection toward the under-sampled, high-entropy corners of materials space that large databases tend to miss by default.

The Materials Genome Initiative, launched by the U.S. government in 2011, bet that shared, ever-growing databases would compress the decades-long materials-to-market timeline. Their work served as a reminder that the bet only pays off if the growing database is well-calibrated.

Materials discovery and automated characterization

A 2025 review co-led by Chen, James Rondinelli, and MIT’s Elsa Olivetti describes a coordinated approach to materials discovery. It uses AI to search decades of published research for instructions on how to make materials, then combines that information with computer simulations and data-driven methods to identify the most promising materials to create and test. The conviction that AI should help researchers analyze materials, not just simulate or design them, drove two additional directions.

One direction developed tools that automatically examine microscope images, identify different features in a material, and sort those features into categories. At the atomic scale, automated crystal system classification from electron diffraction patterns was enabled using multiview opinion fusion machine learning. Another direction introduced UPNet, a supervised AI model that analyzes a material’s atomic structure, predicts its likely properties, and indicates how confident it is in those predictions. By identifying which predictions are reliable and where more information is needed, UPNet can reduce the number of costly physics-based calculations researchers must perform.

AI for energy materials, explicitly

Edward Sargent, a leading energy-materials chemist and director of the Trienens Institute, was among the authors of the paper introducing UPNet, making the study a fitting bridge to Chen's most direct energy-materials contribution. In a 2024 study with battery expert Arumugam Manthiram and other collaborators, Chen’s team developed an AI tool to predict the properties of battery electrolytes—the mixtures of solvents and salts that carry electrical charge inside batteries. Because these ingredients can be combined in countless ways, finding the best mixture through laboratory testing alone is extremely difficult. The tool helps researchers evaluate possible combinations more systematically and identify promising ones faster. This research builds on Chen’s earlier federally funded work using AI-guided methods to improve thin-film solar cells and other energy technologies.

Together, Chen’s investigations help form the framework that supports the Trienens Institute’s AI4Energy enterprise. Elected to the National Academy of Engineering in 2019, the American Academy of Arts & Sciences in 2024, and named a 2023 NSF BRITE Fellow for "AI-Enabled Discovery and Design of Programmable Material Systems," Chen remains one of the field's most active bridges between statistics, machine learning, and materials physics.

Selected Publications

Zhang, Y., Tao, S., Chen, W., & Apley, D. W. (2020). A latent variable approach to Gaussian process modeling with qualitative and quantitative factors. Technometrics, 62(3), 291–302. https://www.tandfonline.com/doi/full/10.1080/00401706.2019.1638834

Wang, Y., Iyer, A., Chen, W., & Rondinelli, J. M. (2020). Featureless adaptive optimization accelerates functional electronic materials design. Applied Physics Reviews, 7(4), 041403. https://pubs.aip.org/aip/apr/article/7/4/041403/831969/Featureless-adaptive-optimization-accelerates

Yerramilli, S., Iyer, A., Chen, W., & Apley, D. W. (2023). Fully Bayesian inference for latent variable Gaussian process models. SIAM/ASA Journal on Uncertainty Quantification, 11(4), 1357–1381. https://epubs.siam.org/doi/10.1137/22M1525600

Wang, L., Yerramilli, S., Iyer, A., Apley, D., Zhu, P., & Chen, W. (2022). Scalable Gaussian processes for data-driven design using big data with categorical factors. Journal of Mechanical Design, 144(2), 021703. https://asmedigitalcollection.asme.org/mechanicaldesign/article/144/2/021703/1116016/Scalable-Gaussian-Processes-for-Data-Driven-Design

Zhang, H., Chen, W. W., Rondinelli, J. M., & Chen, W. (2023). ET-AL: Entropy-targeted active learning for bias mitigation in materials data. Applied Physics Reviews, 10(2), 021403. https://pubs.aip.org/aip/apr/article/10/2/021403/2877876/ET-AL-Entropy-targeted-active-learning-for-bias

Zhang, H., Georgescu, A. B., Yerramilli, S., Karpovich, C., Apley, D. W., Olivetti, E. A., Rondinelli, J. M., & Chen, W. (2025). Emerging microelectronic materials by design: Navigating combinatorial design space with scarce and dispersed data. Accounts of Materials Research. https://pubs.acs.org/amrcda/article/6/6/730/3751204/Emerging-Microelectronic-Materials-by-Design

Zhang, K., Apley, D. W., Chen, W., Liu, W. K., & Brinson, L. C. (2025). A framework for supervised and unsupervised segmentation and classification of materials microstructure images. Acta Materialia, 301, 121588. https://papers.ssrn.com/sol3/papers.cfm?abstract_id=5175316

Chen, J., Zhang, H., Wahl, C., Liu, W., Mirkin, C.A., Dravid, V.P., Apley, D. W., and Chen, W., "Automated crystal system identification from electron diffraction patterns using multiview opinion fusion machine learning", PNAS, 120(46): E2309240120. https://www.pnas.org/doi/10.1073/pnas.2309240120

Chen, J., Ou, P., Chang, Y., Zhang, H., Li, X., Sargent, E. H., & Chen, W. (2026). Materials discovery using uncertainty-aware constrained Bayesian optimization with representation learning of high-dimensional inputs. Journal of Mechanical Design, 148(2), 021707. https://asmedigitalcollection.asme.org/mechanicaldesign/article/148/2/021707/1225557/Materials-Discovery-Using-Uncertainty-Aware