Embedded Files
Â
Our System
Our System
FORUM-AI tackles scientific challenges through a seamlessly integrated, end-to-end system. By combining the FORUM-AI Planner, specialized Domain Agents, and a Multimodal Knowledge Graph, we create a continuous loop of planning, execution, and refinement. This architecture drives our validation against highly demanding use cases, from discovering novel high-entropy cathodes to modeling thermal transport in superlattices and interpreting high-quality experimental measurements.
A chain-of-thought reasoning foundational research planner
A chain-of-thought reasoning foundational research planner
Thrust 1 focuses on developing a versatile research planner capable of reasoning across diverse scientific domains. By integrating multi-modal datasets, including synthesis protocols, characterization data, and vast amount of scientific literature, the agentic system uses a chain-of-thought architecture to decompose complex problems into modular, interpret-able steps. Training this planner requires aligning high-level reasoning with seamless agent interaction, a computational challenge we will meet using DOE leadership-class HPC resources and proven exascale distributed frameworks.
Thrust 1 focuses on developing a versatile research planner capable of reasoning across diverse scientific domains. By integrating multi-modal datasets, including synthesis protocols, characterization data, and vast amount of scientific literature, the agentic system uses a chain-of-thought architecture to decompose complex problems into modular, interpret-able steps. Training this planner requires aligning high-level reasoning with seamless agent interaction, a computational challenge we will meet using DOE leadership-class HPC resources and proven exascale distributed frameworks.
Autonomous agentic systems driving multi-scale simulations, performing experimental operations and accessing material databases
Autonomous agentic systems driving multi-scale simulations, performing experimental operations and accessing material databases
Thrust 2 aims to establish a suite of multi-modal autonomous agents to tackle specific materials science tasks using autonomous experimental and computational toolkits. Operating as modular components, these agents can be called upon independently by researchers or seamlessly orchestrated by the foundational planner. In practice, they act as the computational and physical hands of the FORUM-AI system, capable of generating complex crystal structures, running atomistic simulations, interpreting and assessing simulation/experimental outcomes with uncertain quantifications , and directly controlling experimental equipment.
Thrust 2 aims to establish a suite of multi-modal autonomous agents to tackle specific materials science tasks using autonomous experimental and computational toolkits. Operating as modular components, these agents can be called upon independently by researchers or seamlessly orchestrated by the foundational planner. In practice, they act as the computational and physical hands of the FORUM-AI system, capable of generating complex crystal structures, running atomistic simulations, interpreting and assessing simulation/experimental outcomes with uncertain quantifications , and directly controlling experimental equipment.
Physical grounding of machine learning models through multimodal materials knowledge graphs
Physical grounding of machine learning models through multimodal materials knowledge graphs
A significant challenge in developing autonomous agents is ensuring their outputs are both transparent and trustworthy. Because out-of-the-box LLMs lack the specialized context needed to interpret complex experimental data, Thrust 3 focuses on utilizing self-supervised techniques to extract structured relationships from scientific literature, integrating text, tables, and figures to support accurate, data-driven materials science discovery. To further quantify uncertainty and build user trust, we are deploying advanced validation methods, including Retrieval Augmented Generation (RAG) for fact checking and specialized reflection agents to provide researchers with precise confidence metrics for every model prediction.
A significant challenge in developing autonomous agents is ensuring their outputs are both transparent and trustworthy. Because out-of-the-box LLMs lack the specialized context needed to interpret complex experimental data, Thrust 3 focuses on utilizing self-supervised techniques to extract structured relationships from scientific literature, integrating text, tables, and figures to support accurate, data-driven materials science discovery. To further quantify uncertainty and build user trust, we are deploying advanced validation methods, including Retrieval Augmented Generation (RAG) for fact checking and specialized reflection agents to provide researchers with precise confidence metrics for every model prediction.
Page updated
Report abuse