Multi-Agent AI Enables Emergent Cognition and Real-Time Knowledge Synthesis in Science and Engineering

In the rapidly evolving landscape of artificial intelligence, we stand at the precipice of a new era in scientific and engineering discovery. The question that has long tantalized researchers is: Can, and if yes, how, AI models solve complex scientific tasks in an autonomous, self-improving manner? The answer, we believe, lies in agentic modeling, specifically through multi-agent, multimodal AI frameworks. Such agentic models can be implemented based on theory and simulation but also seamlessly integrate withe experimental platforms, such as self-driving labs.

Multi-agent AI systems draw inspiration from the evolutionary principles of biological systems, which have optimized, and re-used, relatively simple components into highly functional, complex structures over billions of years. The AI model building blocks in agentic systems leverage similar strategies to create sophisticated, autonomous systems capable of solving complex scientific and engineering problems. Thereby, out of many relatively simple elements, AI systems emerge with higher capabilities and complexity, leveraging similar strategies to create sophisticated, autonomous systems.

In several recent papers we reported examples for such approaches for a variety of tasks such as protein modeling, analysis and design, molecular mechanics, as well as solving mechanics problems. These models represent a significant evolution in the established paradigm, autonomously planning experiments, soliciting new data from simulations, and reasoning over intermediate results, images and graphs, visualizations of molecular dynamics results, fracture mechanisms, and others, to drive scientific discovery forward.

What Current AI Models are Missing

The landscape of artificial intelligence, while advancing at a breathtaking pace, still falls short of the dynamic, adaptive intelligence required for genuine scientific discovery. Current AI models, despite their impressive capabilities, remain tethered to static paradigms that limit their true potential.

At the core of the limitation of many existing AI models lies the passive nature of their knowledge acquisition. These models, while capable of impressive feats of inference and even generating novel ideas within their training distribution, still lack the crucial ability to autonomously design and execute empirical tests of their hypotheses. They can propose intriguing possibilities, but cannot independently validate them against physical reality.

The baked-in knowledge of these models, while extensive, is fundamentally constrained by the temporal boundaries of their training data. This creates an artificial ceiling on their ability to adapt to new discoveries or paradigm shifts in scientific understanding. The absence of real-world feedback mechanisms further exacerbates this limitation, divorcing these models from the empirical foundation that underpins scientific progress.

Perhaps most critically, current AI models lack the capacity for causal reasoning and self-reflection. They excel at correlation but falter at causation, a distinction that lies at the heart of scientific inquiry. Their inability to recognize the boundaries of their own knowledge – to know what they don't know – hampers their capacity for targeted learning and strategic exploration of unknown domains.

The mono-modal nature of most AI systems further constrains their ability to integrate diverse forms of information, a skill that human scientists employ effortlessly. This limitation, coupled with their inability to generate new data through experimentation or simulation, creates a fundamental barrier to their engagement in true scientific discovery.

In essence, what current AI models lack is not just specific capabilities, but the very essence of scientific thinking – the ability to question, explore, and reshape understanding based on new evidence. It is this gap that multi-agent systems aim to bridge, promising a new era where AI can not only assist in scientific endeavors but actively drive the frontiers of human knowledge.

The Philosophy Behind Multi-Agent AI Systems

The concept of multi-agent AI systems is rooted in a philosophical shift in how we approach artificial intelligence and problem-solving. This approach draws inspiration from human cognitive processes, teamwork, and social structures, recognizing that complex problems often require diverse perspectives and specialized knowledge, capabilities and information.

1. Distributed and Adversarial Intelligence: Multi-agent systems embody the idea that intelligence can be distributed across multiple entities, mimicking how human societies solve complex problems through collaboration and competition.
2. Emergent Behavior: These systems operate on the principle that the collective intelligence of multiple agents can surpass the capabilities of any single agent, leading to emergent behaviors and solutions.
3. Specialization and Collaboration: Just as human societies benefit from specialized roles, multi-agent AI systems leverage the power of focused expertise combined with collaborative problem-solving.
4. Adaptability and Flexibility: By design, these systems can adapt to new information and changing conditions, reflecting the dynamic nature of scientific inquiry. They will error correct and self-learn from feedback from simulations, data, collected information and other agents.
5. Ethical Considerations: The development of multi-agent AI systems raises important philosophical questions about decision-making, accountability, and the nature of intelligence itself.

This philosophical underpinning guides the design and implementation of agentic models, enabling them to tackle complex scientific challenges in novel ways.

The Power of Multi-Agent AI

At its core, multi-agent AI is a system where multiple intelligent agents interact to solve complex problems. Think of it as a virtual team of experts, each with their own specialized knowledge and capabilities, working in concert to tackle challenges that would be insurmountable for a single agent.

We have designed teams of AI agents, each with a specific role:

1. Administrator: Represents the human user, posing questions and approving plans.
2. Planner: Responsible for developing strategies to solve tasks, suggesting appropriate functions and tools.
3. Assistant/Engineer: Executes the functions and tools as directed by the Planner, writes and debugs code.
4. Scientist: Formulates problems using domain-specific knowledge (e.g., mechanics principles for mechanics problems).
5. Executor: Runs codes, simulations and returns outcomes, including new data.
6. Critic: Provides feedback on plans and results, catching potential errors and suggesting improvements.
7. Chat Manager: Orchestrates the interaction between all agents, ensuring smooth collaboration and interactions towards problem solving.

This division of labor allows for a more nuanced and comprehensive approach to problem-solving, mirroring the collaborative nature of human scientific teams.

The Inner Workings

The approaches are modular and allow to seamlessly integrate various physics-based computational and AI capabilities:

1. Flexible Platform: At their heart, these frameworks utilize multiple AI agents with distinct capabilities. These range from general-purpose intelligence powered by large language models (in our case, GPT-4) to specialized models for tasks like protein folding or finite element analysis.
2. Modular Structure: Inspired by biological systems, our frameworks combine small, specialized subunits into an integrated model. This approach allows for emergent scientific and engineering abilities that surpass the sum of its parts.
3. Autonomous Planning: Unlike traditional AI systems that rely on predefined workflows, our agents can autonomously plan their approach to a problem. They can break down complex tasks into manageable steps, decide which tools or simulations to use, and even revise their plans based on intermediate results.
4. Collaborative AI Agents: The true power of these frameworks lies in the dynamic collaboration between their agents. They can discuss strategies, critique each other's work, and collectively reason about the best approach to a problem.
5. Acquisition and Integration of New Data: Perhaps most excitingly, the frameworks we developed can recognize when they need more information and can solicit new data from simulations, write code to do acquire and analyze new data, or exploit other sources such as scientific papers, search engines, and others. This ability to "know what they don't know" and seek out new information is a crucial step towards truly autonomous scientific discovery.

Autonomous Data Generation and Retrieval

One of the critical aspects of multi-agent AI systems is their ability to generate, collect, and analyze their own data. This capability represents a significant leap forward in autonomous scientific discovery:

1. Code Generation for Data Creation: These systems can write their own code to generate data through simulations or computational models. For instance, MechAgents demonstrated the ability to write complex finite element analysis code to solve elasticity problems. The ability to reason over results (images, numbers, graphs, etc.) and relate them with existing data or earlier results provides opportunities for relational feedback and enhanced reasoning.
2. Running Physics Models: Once the code is generated, the systems can execute these physics-based models to produce new data. This includes running simulations for protein folding in ProtAgents, solving hyperelasticity problems in MechAgents, or executing first-principles simulations.
3. Data Retrieval from Literature, Patents, Documents and more: The agents are capable of extracting relevant information from scientific papers and databases. This allows them to incorporate existing knowledge into their problem-solving processes.
4. Result Analysis and Iteration: After generating or retrieving data, the systems can analyze the results, identify patterns or anomalies, and use this information to refine their approach or generate new hypotheses.
5. Continuous Learning: Through this process of data generation, retrieval, and analysis, these multi-agent systems can continuously expand their knowledge base and improve their problem-solving capabilities. This happens through in-context learning and can be enhanced through dynamic model weight updating or reinforcement mechanisms.

This autonomous data handling capability allows these AI systems to operate with a level of independence previously unseen in scientific AI applications, potentially accelerating the pace of discovery across various scientific domains.

Impacts and Possibilities

The potential impacts of this technology extend far beyond protein design and mechanics:

1. Comprehensive Problem-Solving: By incorporating out-of-domain knowledge and enabling thorough data analysis, these frameworks can tackle problems that traditional models struggle with.
2. Versatility: Our experiments demonstrate the systems' ability to design new proteins, analyze complex structures, solve elasticity problems, and obtain first-principles data via simulations. This versatility makes them applicable to a wide range of scientific disciplines.
3. Targeted Design and Discovery: ProtAgents, for instance, allows for the automated design of de novo proteins with specific mechanical properties. MechAgents can solve complex elasticity problems with various boundary conditions and material properties. These capabilities could revolutionize materials engineering, enabling the rapid development of new materials with precisely tuned characteristics.
4. Bridging the Gap Between AI and Scientific Simulation: By integrating AI agents with physics-based simulations, these frameworks represent a significant step towards closing the loop between AI prediction and scientific validation. This provides particularly interesting avenues to enhance technologies such as digital twins.
5. Accelerating Scientific Discovery: The autonomous nature of these systems could dramatically speed up the scientific process, allowing for rapid exploration of vast design spaces that would be impractical for human researchers to navigate manually.

Example Results

ProtAgents: Solving protein analysis and design problems

1. Knowledge Retrieval and Multi-Step Analysis: ProtAgents successfully retrieved information about proteins whose mechanical properties had been studied experimentally and performed complex multi-step analyses on the identified proteins.
2. De Novo Protein Design and Analysis: The system designed new proteins with specified lengths, analyzed their secondary structures, computed vibrational frequencies, and performed folding simulations.
3. Structure-Property Relationships: ProtAgents investigated the relationship between protein secondary structure and mechanical properties, designing and analyzing proteins with varying structural compositions.

MechAgents: Solving mechanical engineering problems

1. Linear Elasticity Problems: The system successfully solved a series of increasingly complex linear elasticity problems, including changing boundary conditions, refining meshes, and adding geometric features like circular holes.
2. Hyperelasticity: MechAgents demonstrated the ability to transition from linear elasticity to hyperelasticity problems, correctly implementing a compressible Neo-Hookean model with finite deformation.
3. Complex Geometry and Post-Processing: The system handled a challenging problem involving a plate with a circular hole, correctly implementing the geometry, solving the elasticity problem, and computing the total force on an edge.

Future Possibilities for AI-Driven Scientific Discovery

The introduction of ProtAgents and MechAgents represents more than just new tools for protein design and mechanics problem-solving; it heralds a new paradigm in AI-driven scientific discovery. By combining the power of large language models, specialized scientific tools, and physics-based simulations within flexible, multi-agent frameworks, we're opening up new possibilities for solving complex problems across a wide range of scientific disciplines.

Looking ahead, we envision several exciting directions for this technology:

1. Expansion to Other Domains: While our initial focus has been on protein design, molecular modeling, and mechanics, the principles underlying these frameworks could be applied to other areas of materials science, drug discovery, or even fields like climate modeling or astrophysics.
2. Integration with Robotic Systems: Imagine a future where AI agents not only plan experiments but also control robotic systems to carry them out, creating a closed-loop system for autonomous scientific discovery. Self-driving labs are already a reality, and can be supercharged with multi-agent AI systems.
3. Collaborative Human-AI Research: These frameworks could serve as intelligent assistants to human researchers, suggesting experiments, analyzing results, and even generating hypotheses for further investigation.
4. Accelerated Materials Discovery: By rapidly exploring vast design spaces and predicting properties, systems like ProtAgents and MechAgents could dramatically speed up the discovery of new materials with tailored properties.
5. Enhanced Interpretability: As these systems become more complex, developing ways to interpret and explain their decision-making processes will be crucial. This could lead to new insights into the scientific process itself. The ability to relate with existing literature, images, graphs and other real-world feedback allows for unparalleled cognition across scales, from nano to macro.

Concluding Thoughts

The development of agentic models such as ProtAgents and MechAgents represent a step forward in our quest for autonomous scientific discovery through AI. By leveraging the power of multi-agent AI, we are creating systems that can not only assist human researchers but also drive scientific inquiry and engineering problem solving forward, on their own.

The ability to collaboratively work with such agentic systems, through human-in-the-loop strategies and embodied intelligence, is particularly appealing as a workflow that embeds naturally with existing technologies and team organization.

There are various ways to construct multi-agent systems, such as writing custom code, graph-based methods, and powerful frameworks such as AutoGen that we used in several of our studies. AutoGen is an open-source framework that provides a robust foundation for building and orchestrating AI agents that offers a flexible and scalable infrastructure that enables the creation of diverse, specialized agents and facilitates their seamless interaction. Its modular architecture allows for the integration of various AI models, including multimodal large language models like GPT-4, GPT-4o, Claude 3.5, and custom tools tailored for specific scientific domains.

As we continue to refine and expand this technology, we anticipate transformative advancements that will change how we approach complex scientific and engineering challenges. The future of scientific discovery is collaborative, autonomous, and AI-driven, and builds on many paradigms biological systems have developed and exploited over billions of years, to create highly functional systems from simple constituents.

We hope that many other researchers, engineers, and innovators will explore the possibilities in this emerging field so that we can push the boundaries of what is possible in AI-driven engineering and scientific discovery.

References

• Buehler, M. J. (2024). Generative retrieval-augmented ontologic graph and multiagent strategies for interpretive large language model-based materials design. ACS Engineering Au, 4(2), 241–277. https://doi.org/10.1021/acsengineeringau.3c00058
• Ni, B., & Buehler, M. J. (2024). MechAgents: Large language model multi-agent collaborations can solve mechanics problems, generate new data, and integrate knowledge. Extreme Mechanics Letters, 67, 102131. https://doi.org/10.1016/j.eml.2024.102131
• Ghafarollahi, A., & Buehler, M. J. (2024). ProtAgents: Protein discovery via large language model multi-agent collaborations combining physics and machine learning. Extreme Mechanics Letters, 67, 102131. https://doi.org/10.1016/j.eml.2024.10213
• Wu, Q., Bansal, G., Zhang, J., Wu, Y., Li, B., Zhu, E., Jiang, L., Zhang, X., Zhang, S., Liu, J., Awadallah, A. H., White, R. W., Burger, D., & Wang, C. (2023). AutoGen: Enabling Next-Gen LLM Applications via Multi-Agent Conversation. arXiv. https://doi.org/10.48550/arXiv.2308.08155