Public Abstract

DE-SC0025801: Harnessing Nonnegative Matrix Factorization for Advanced Computational Materials Modeling

Award Status: Active

Institution: Lehigh University, Bethlehem, PA
UEI: E13MDBKHLDB5
DUNS: 808264444

Most Recent Award Date: 01/17/2025
Number of Support Periods: 1
PM: Spotz, William

Current Budget Period: 12/01/2024 - 11/30/2025
Current Project Period: 12/01/2024 - 11/30/2027
PI: Ekuma, Chinedu

Supplement Budget Period: N/A

Public Abstract

Harnessing Nonnegative Matrix Factorization for Advanced Computational Materials Modeling Chinedu Ekuma, Lehigh University, Bethlehem, PA 18015 (Principal Investigator) Lifang He, Lehigh University, Bethlehem, PA 18015 (Co-Investigator) Akwum Onwunta, Lehigh University, Bethlehem, PA 18015 (Co-Investigator) Bao Wang, University of Utah, Salt Lake City, UT 84112 (Co-Investigator) As conventional materials near performance limits, developing novel materials with enhanced functionalities is essential for advancing next-generation devices. The search for these materials – especially low-dimensional crystal structures – presents computational challenges due to their complex architectures and the need for high-fidelity simulations, emphasizing the need for innovative approaches in materials discovery. This project aims to address these challenges by creating a new class of scientific machine learning (SciML) algorithms tailored to interpret high-dimensional materials data. By leveraging deep learning (DL)-assisted non-negative matrix factorization (NMF) and diffusion models (DMs), the project establishes a scalable framework for capturing intricate structure-property relationships, enabling precise tuning of electronic and material property predictions. This approach will drive advancements in optoelectronic and quantum materials for applications across energy, environmental sustainability, and beyond. The project’s primary objective is to develop physics-informed NMF models that incorporate structural symmetries and physical principles, providing interpretable representations that capture fundamental material behaviors. This objective is achieved through integrative aims: developing structure-aware NMF models grounded in time-frequency features and physical laws to enhance interpretability (Aim 1); creating scalable machine learning models based on these NMF foundations to improve property predictions and address data scarcity through randomized deep NMF and Bayesian inference (Aim 2); integrating NMF with diffusion models to process complex datasets and uncover correlations in high-throughput density functional theory (DFT) simulations (Aim 3); and establishing an open-source platform that consolidates these tools, accessible across diverse computing environments, to facilitate broad scientific adoption (Aim 4). Together, these aims build a foundation for data-driven discovery and efficient design of novel materials with applications across energy, sustainability, and advanced manufacturing. This research supports the DOE mission by creating efficient, scalable, and interpretable tools that empower scientific discovery. By incorporating theoretical foundations and uncertainty quantification, the project transforms “black-box” models into transparent, reliable tools for diverse applications, significantly impacting fields such as energy, environmental sustainability, climate modeling, healthcare analytics, and advanced manufacturing. The open-source nature of the platform democratizes access to advanced tools, enabling researchers across disciplines to tackle complex scientific challenges with confidence. This project also emphasizes education and outreach initiatives. The PIs will promote diversity and inclusivity by mentoring students from underrepresented backgrounds in STEM, fostering a diverse future workforce skilled in computational materials science. Students will gain hands-on experience in computational physics, data science, and materials research, cultivating critical algorithmic thinking and simulation expertise.

Harnessing Nonnegative Matrix Factorization for Advanced Computational Materials Modeling

Chinedu Ekuma, Lehigh University, Bethlehem, PA 18015 (Principal Investigator)

Lifang He, Lehigh University, Bethlehem, PA 18015 (Co-Investigator)

Akwum Onwunta, Lehigh University, Bethlehem, PA 18015 (Co-Investigator)

Bao Wang, University of Utah, Salt Lake City, UT 84112 (Co-Investigator)

As conventional materials near performance limits, developing novel materials with enhanced functionalities is essential for advancing next-generation devices. The search for these materials – especially low-dimensional crystal structures – presents computational challenges due to their complex architectures and the need for high-fidelity simulations, emphasizing the need for innovative approaches in materials discovery. This project aims to address these challenges by creating a new class of scientific machine learning (SciML) algorithms tailored to interpret high-dimensional materials data. By leveraging deep learning (DL)-assisted non-negative matrix factorization (NMF) and diffusion models (DMs), the project establishes a scalable framework for capturing intricate structure-property relationships, enabling precise tuning of electronic and material property predictions. This approach will drive advancements in optoelectronic and quantum materials for applications across energy, environmental sustainability, and beyond. The project’s primary objective is to develop physics-informed NMF models that incorporate structural symmetries and physical principles, providing interpretable representations that capture fundamental material behaviors. This objective is achieved through integrative aims: developing structure-aware NMF models grounded in time-frequency features and physical laws to enhance interpretability (Aim 1); creating scalable machine learning models based on these NMF foundations to improve property predictions and address data scarcity through randomized deep NMF and Bayesian inference (Aim 2); integrating NMF with diffusion models to process complex datasets and uncover correlations in high-throughput density functional theory (DFT) simulations (Aim 3); and establishing an open-source platform that consolidates these tools, accessible across diverse computing environments, to facilitate broad scientific adoption (Aim 4). Together, these aims build a foundation for data-driven discovery and efficient design of novel materials with applications across energy, sustainability, and advanced manufacturing. This research supports the DOE mission by creating efficient, scalable, and interpretable tools that empower scientific discovery. By incorporating theoretical foundations and uncertainty quantification, the project transforms “black-box” models into transparent, reliable tools for diverse applications, significantly impacting fields such as energy, environmental sustainability, climate modeling, healthcare analytics, and advanced manufacturing. The open-source nature of the platform democratizes access to advanced tools, enabling researchers across disciplines to tackle complex scientific challenges with confidence.

This project also emphasizes education and outreach initiatives. The PIs will promote diversity and inclusivity by mentoring students from underrepresented backgrounds in STEM, fostering a diverse future workforce skilled in computational materials science. Students will gain hands-on experience in computational physics, data science, and materials research, cultivating critical algorithmic thinking and simulation expertise.