In a world where scientific research and the development of new materials often face setbacks due to time, cost, and experimental constraints, researchers from MIT have launched a revolutionary system that promises to change the way scientific hypotheses are tested and confirmed. They named it CRESt — “Copilot for Real-world Experimental Scientists”. This innovative system combines artificial intelligence, robotics, machine learning, and human intuition into a unique experimental ecosystem that independently plans, conducts, monitors, and adjusts material experiments.

Behind the Idea: Challenges of Traditional Methods in Materials Science

Current methods for discovering and optimizing new materials often depend on manual experiment design, iterative modifications, and a great deal of trial and error. Researchers must design protocols, synthesize samples, conduct measurements, analyze results, and then decide which direction to consider next. This approach wastes time and resources—especially when the number of variables escalates rapidly.

Until now, many attempts to apply machine learning in this field have relied on simple models with a single data stream, such as comparing compositions and resulting properties, or using active learning within the framework of Bayesian optimization. However, these model-based approaches are often limited—they address only a small fraction of the complexity in the relationships between chemical compositions, processing parameters, microstructures, and experimental conditions.

Furthermore, the reproducibility of experiments often suffers due to nuanced differences in sample preparation, equipment drift, unforeseen deviations, and inconsistencies. To achieve true autonomy in the laboratory, a system must be able to not only perform experiments but also monitor them, detect errors, and adapt in real time.

How CRESt Works: Integration of Multimodal Data and Robotic Execution

The main advantage of the CRESt system is its ability to combine various sources of knowledge—written papers and literature, chemical compositions, microstructural images (e.g., SEM, XRD), process parameters, and measurement results—and merge them into a single, AI-driven strategy. The system uses a large multimodal model (LMM) that learns from all these modalities and makes decisions about further experiments.

When a researcher gives instructions in the user interface (textually or vocally), CRESt analyzes the problem, incorporates literature and historical data, formulates experimental proposals, and initiates robots to execute them. During the experiment, the system uses built-in cameras and image processing models to monitor performance: it can detect a pipette shift, an irregular sample shape, or an unexpected change in the system and automatically suggest adjustments.

The system's robotic infrastructure includes liquid-handling equipment, devices for rapid material synthesis (e.g., "carbothermal shock" systems), automated electrochemical stations for performance testing, and microscopic instruments for sample characterization. The system also controls pumps, ventilation systems, and valve components, often remotely or automatically.

How does CRESt select the next experiments? First, the system uses embeddings based on literature and a database to represent potential chemical recipes in a high-dimensional space. Then, analytical methods such as principal component analysis (PCA) reduce the dimensionality into this representation space, retaining the most important variables. On this reduced space, Bayesian optimization is performed to select the next promising state of experiments. The obtained results (chemistry + parameters + performance) are fed back into the model, allowing the system to learn and update its decisions.

If the system detects an irregularity in the experiment, it uses visual models and domain information to propose corrective measures. In this way, it not only learns from the results but also reacts to real events in the laboratory.

Demonstration Results: A New Generation Catalyst for Direct Formate Fuel Cells

The team of researchers at MIT conducted an experimental cycle using CRESt systematically, testing over 900 chemical formulations and performing about 3,500 electrochemical measurements over a period of several months. This experiment led to the discovery of a catalyst composed of eight elements, which achieved a record power density in a direct formate fuel cell (using formate as fuel). The system managed to achieve a power-per-dollar ratio 9.3 times better compared to pure palladium catalysts, while using only a quarter of the precious metals compared to previous standards.

It is important to note that it was not predetermined for the system to use the initial elements—CRESt itself explored combinations, included different elements, and iteratively optimized the results. This led to a reduction in costs and an increase in efficiency, bringing the goal of sustainable and economically viable fuel cell technologies closer.

Advantages and Challenges: Where CRESt Shines and Where It Needs Improvement

The advantages of this approach are obvious:

• faster iterations in experimentation;


• less need for manual interventions;


• use of diverse data in a single coherent strategy;


• better reproducibility thanks to real-time monitoring;


• democratization—researchers without advanced programming knowledge can also use the system via natural language.

However, there are also obstacles to overcome. Systems like CRESt require very precise hardware integrations and a robust software framework to withstand the risks of experimental variations. Managing noise, uncertainties, and random errors in the laboratory remains a challenge. A review of the literature on autonomous laboratories emphasizes that it is crucial to incorporate a design that enables reproducibility, interoperability, and repeatable protocols into systems that experiment autonomously.

Furthermore, relying on a large number of automated experiments can create a "data overload" problem—the system must be able to filter and select the most relevant information. In this sense, methodological papers from the field of active learning show that the choice of acquisition functions, the balance between exploitation and exploration, and the incorporation of physical laws into models (so-called scientific machine learning) can significantly affect the system's efficiency.

A Look into the Future and Implications for Scientific Practice

The CRESt system shows that it is possible to build autonomous laboratories that not only operate on rigid scripts but also learn, adapt, and communicate in human language. This approach opens the door to "scientific mass production"—parallel experiments, data sharing in the cloud, and collaboration among laboratories worldwide.

Improvements already in development include the integration of multiple artificial intelligence agents within CRESt, which "talk" to each other to improve image processing and decision-making in material analysis. Such multi-agent approaches contribute to better accuracy in identifying phases and structures in materials.

Expanding such systems from the laboratory scale to industrial processes, automated pilot plants, and mass production will require additional steps: standardization of protocols, modularity of lab components, an open data exchange format, and interoperability between different autonomous laboratories.

Ultimately, CRESt does not aim to replace the human scientist—but to empower them. The system explains its hypotheses and processes in natural language and enables collaboration between human and machine in solving the most complex problems. While machines optimize and automate, humans remain the creators, thinking critically and directing the scientific flow toward an unknown horizon.