MIT’s CRESt AI System Combines Multimodal Data and Robotics to Accelerate Materials Discovery

A new AI-driven platform called Copilot for Real-world Experimental Scientists (CRESt), developed by MIT researchers, is transforming materials discovery by combining machine learning, robotics, and multimodal data to autonomously design and conduct experiments. Unlike traditional AI models that rely on limited data types, CRESt integrates diverse information sources—including scientific literature, chemical compositions, microstructural images, and real-time experimental feedback—enabling it to make more informed decisions, much like a human research team. The system uses a network of robotic tools, including a liquid-handling robot, a carbothermal shock system for rapid material synthesis, automated electrochemical testers, and advanced characterization equipment like electron and optical microscopes. These tools allow CRESt to prepare, test, and analyze materials with high throughput. Researchers interact with the system using natural language, without needing to write code, making it accessible and intuitive. At the core of CRESt is a multimodal approach that enhances active learning. The system first draws on scientific literature and databases to create rich representations of potential material recipes. It then applies principal component analysis to reduce the search space to a manageable set of key variables, enabling more efficient Bayesian optimization. After each experiment, the results—along with visual data from cameras and microscopes—are fed back into large language models to update the knowledge base and refine future experiments. One major challenge in materials science is reproducibility. Small variations in mixing, processing, or equipment can drastically affect outcomes. To address this, CRESt uses computer vision and vision-language models to monitor experiments in real time. It can detect anomalies—like a misaligned sample or a pipette error—and suggest fixes through text or voice, acting as a real-time assistant. In a key demonstration, CRESt explored over 900 material chemistries and conducted 3,500 electrochemical tests to develop a new catalyst for direct formate fuel cells. The resulting material, composed of eight elements including inexpensive non-precious metals, achieved a 9.3-fold improvement in power density per dollar compared to pure palladium. Remarkably, it delivered record performance while using only a quarter of the precious metals required in prior designs. The research team, led by Ju Li, emphasizes that CRESt is not a replacement for human scientists but a powerful collaborator. While the system handles repetitive tasks, data analysis, and hypothesis generation, human researchers remain essential for oversight, interpretation, and complex decision-making. The platform’s ability to explain its reasoning in natural language helps build trust and transparency. CRESt represents a major step toward self-driving laboratories—autonomous systems capable of continuous learning and adaptation. By merging AI, robotics, and human insight, it accelerates the pace of discovery in critical areas like clean energy and advanced materials, offering a scalable solution to long-standing scientific challenges.