FMAL Research

Collaborators: Weist (Chemistry), Wang (UCLA), Coley (MIT)
Funding: NSF C-CAS

Although existing foundation models have shown promise in answering general chemistry queries, our studies reveal several persistent limitations. These include difficulty interpreting SMILES representations, hindering their ability to complete real chemistry tasks, and a lack of reasoning capabilities in molecular structure elucidation.

As a result, adapting LLMs to chemistry remains a significant challenge, particularly due to the severe data limitations inherent to these complex, domain-specific problems.

We address these challenges by building chemistry-specialized foundation models designed to assist chemists in solving real-world tasks, such as reaction prediction, structure interpretation, and property inference, with enhanced reasoning and domain-aligned representations.

Collaborators: Georgina Curto Rex (Senior AI researcher and team lead, United Nations University Institute in Macau)

By integrating foundation models into agent-based simulations, we create virtual environments where autonomous agents, representing diverse individuals and communities, interact in ways that reflect the real world.

Empowered by the abilities of foundation models in contextual understanding, reasoning, opinion formation and expression through text generation, these agents will be designed to mirror human-like behaviors, including decision-making, communication, and adaptation to socio-economic changes.

Collaborators: Chen (IBM), Gao (IBM)
Funders: ND-IBM Tech Ethics Lab

Foundation models increasingly assist in tasks ranging from procedural guidance to autonomous experiment orchestration. However, their advanced capabilities also introduce significant risks when hallucinations, delayed hazard detection, or protocol misinterpretations intersect with physical lab operations.

Failures in adhering to safety protocols or prioritizing research objectives can escalate into critical incidents. Our study focuses on understanding and enhancing the reliability of foundation models in safety-critical tasks.

Collaborators: Luo (ND AME), Guo (ND ChemE)
Funded by NSF CBET

This project aims to combine molecular simulations, machine learning, and experiments to develop amorphous polymers with high TC (>1.5 W/mK) and understand the underlying molecular features dictating thermal transport.

More info: NSF Award

Collaborators: Errington, Shakhlo (COS), Wu (ODU), Rajtmajer (PSU)
Funders: Center for Open Science

This project aims to develop a multimodal foundation model agent that automatically generates and validates scientific hypotheses. The agent is expected to be a helpful assistant for researchers.

Collaborator: Wu (ND Math)

This project aims to develop an efficient system for evaluation of AI capabilities with respect to professional mathematics such as partial differential equations (Navier Stokes), number theory (Riemann Hypothesis), and computational complexity.

Evaluation is based on correctly positing the lemma, formalizing, proving, and stating the proof in natural language.

Collaborators: Wang, Huang (UTK EEPS)

This project aims to create general-purpose AI systems that can accelerate scientific discovery by learning from large, diverse datasets across planetary science.

Geoscience foundation models aim to scale up the reasoning, discovery, and exploration capabilities of planetary scientists by acting as versatile, data-driven scientific assistants.

Collaborators: Zhang, Haeffel (ND Psychology), Sullivan (IBM-ND Tech Ethics)

This project aims to develop new model capabilities that would lead to more competent models when applied to mental health.

It also aims to develop new methods and datasets to evaluate and benchmark model performance when applied to mental health measurement.