Engineers have invented AI frameworks designed to discover and refine research hypotheses aimed at enhancing biologically inspired materials.
Formulating an original and viable research hypothesis is essential for any scientist, although it can also take a lot of time. Many first-year PhD students devote significant time just figuring out what to investigate in their research. Imagine if artificial intelligence could assist with this process?
Researchers at MIT have introduced a method to autonomously create and assess promising research hypotheses across various fields using collaboration between humans and AI. Their latest publication explains how they employed this framework to formulate evidence-based hypotheses that meet unaddressed research needs in biologically inspired materials.
Released in Advanced Materials, the study is co-authored by Alireza Ghafarollahi, a postdoctoral fellow in the Laboratory for Atomistic and Molecular Mechanics (LAMM), alongside Markus Buehler, the Jerry McAfee Professor in Engineering at MIT, who leads LAMM.
The framework, referred to as SciAgents, comprises several AI agents, each with distinct capabilities and data access. They utilize “graph reasoning,” wherein AI models use a knowledge graph that organizes and defines the relationships among various scientific concepts. This multi-agent approach is akin to how biological systems arrange themselves into groups of fundamental building blocks. Buehler highlights that this “divide and conquer” strategy is a key principle in biology, ranging from materials to insect swarms to entire civilizations—scenarios where the overall intelligence far exceeds the combined abilities of individual elements.
“By employing multiple AI agents, we aim to replicate how scientific communities make discoveries,” explains Buehler. “At MIT, this occurs through the collaboration of individuals with different backgrounds, interacting informally in places like coffee shops or the Infinite Corridor. However, this is often coincidental and slow. Our goal is to simulate discovery processes through investigating whether AI can be creative and contribute to new findings.”
Streamlining the process of generating ideas
Recent advancements have shown that large language models (LLMs) possess a commendable capacity to answer questions, summarize information, and complete simple tasks, but they struggle with generating new ideas independently. The MIT team aimed to design a system that empowers AI models to engage in a more nuanced, multi-step process, exceeding the mere retrieval of learned information to create new knowledge.
At the core of their strategy is an ontological knowledge graph, which organizes and connects various scientific concepts. To build these graphs, the researchers input a selection of scientific papers into a generative AI model. In earlier work, Buehler applied category theory—a branch of mathematics—to assist the AI in forming abstractions of scientific concepts as graphs, focused on defining relationships in a manner that other models can analyze using graph reasoning. This approach encourages AI models to develop a more principled understanding of concepts, allowing them to generalize more effectively across different areas.
“This is crucial for us to create AI models centered on science, as scientific theories usually rest on broadly applicable principles rather than mere fact recollection,” Buehler comments. “By training AI to ‘think’ in this way, we can surpass traditional methods and explore more innovative applications of AI.”
In the latest paper, the researchers utilized approximately 1,000 scientific studies on biological materials; however, Buehler adds that knowledge graphs can be compiled from a broader or narrower range of research papers from any discipline.
Once the graph is created, the researchers designed an AI system for scientific discovery, consisting of various models each tasked with specific roles within the system. Many components were derived from OpenAI’s ChatGPT-4 series models and utilized a technique known as in-context learning, where prompts provide contextual clues about the model’s role while allowing for learning from given data.
The individual agents in this architecture collaborate to address a complex problem that none could tackle alone. Their initial assignment is to formulate the research hypothesis. The interactions involving Large Language Models commence after a subgraph is delineated from the knowledge graph, accomplished either randomly or by inputting specific keywords discussed in the papers.
Within this framework, a language model dubbed the “Ontologist” is responsible for defining scientific terms in the papers and exploring their interconnections to enhance the knowledge graph. Following this, a model called “Scientist 1” formulates a research proposal taking into account factors like the potential for unexpected findings and novelty. This proposal discusses anticipated results, the significance of the research, and hypotheses regarding underlying mechanisms. A subsequent model, “Scientist 2,” further develops the idea by proposing specific experimental and simulation methodologies, as well as suggesting enhancements. Ultimately, a “Critic” model evaluates strengths and weaknesses, providing constructive feedback.
“It’s about assembling a team of experts with diverse thinking patterns,” Buehler remarks. “They need to have unique perspectives and skills. The Critic agent is intentionally designed to challenge the others, ensuring that not everyone agrees unconditionally—some may say, ‘There’s a flaw here; can you elaborate?’ This dynamic generates outputs significantly different from what a single model could produce.”
Additional agents in the system are equipped to search existing literature, enabling the system to evaluate feasibility and assess the originality of each idea.
Enhancing the system’s capabilities
To validate their method, Buehler and Ghafarollahi developed a knowledge graph centered on the terms “silk” and “energy intensive.” Within this framework, “Scientist 1” suggested merging silk with pigments derived from dandelions to formulate biomaterials boasting superior optical and mechanical traits. The model predicted that this new material would be significantly stronger than conventional silk and would require less energy for processing.
Scientist 2 then provided recommendations, such as implementing specific molecular dynamic simulation tools to investigate how the proposed materials would behave and noting that a potential application might be as a bioinspired adhesive. The Critic model highlighted numerous advantages of the new material while also addressing potential weaknesses, such as scalability, long-term stability, and environmental concerns related to solvent use, recommending pilot studies for process validation and thorough examinations of material durability.
The researchers also explored other experiments with randomly chosen keywords, generating distinct hypotheses about more efficient biomimetic microfluidic chips, improving the mechanical properties of collagen scaffolds, and investigating the interaction between graphene and amyloid fibrils to produce bioelectronic devices.
“The system generated these innovative, robust ideas based on the knowledge graph,” Ghafarollahi states. “Regarding originality and applicability, the materials appeared to be solid and innovative. In forthcoming studies, we plan to generate thousands, or even tens of thousands, of new research ideas, categorizing them to better understand their generation and potential enhancements.”
In the future, the researchers aspire to integrate new tools for information retrieval and simulation within their frameworks. They can also effortlessly update the foundational models in their systems to incorporate the latest advancements in AI technology.
“Due to the interactive nature of these agents, even a minor enhancement in one model can significantly impact the overall performance and results of the system,” Buehler explains.
Since making a preprint of their approach publicly available with open-source details, the researchers have received outreach from hundreds of individuals interested in utilizing their frameworks across various scientific fields as well as in areas like finance and cybersecurity.
“There’s a multitude of tasks you can accomplish without needing to step into the lab,” Buehler emphasizes. “Ideally, you want to conduct experimental work only toward the end of the process. Labs are costly and time-consuming, so it’s essential to have a system that thoroughly investigates the best ideas, formulates the strongest hypotheses, and accurately forecasts emergent behaviors. Our vision is to simplify this process, allowing users an app interface to integrate diverse ideas or datasets to genuinely challenge the model in making new discoveries.”
