MIT achievement published in the main issue of Nature: In 90 days, an "AI scientist" completed 3,500 electrochemical tests.

The research team led by Ju Li from the Massachusetts Institute of Technology (MIT) published a research paper in the world's top academic journal Nature, presenting a multimodal robotic platform called CRESt (Copilot for Real-world Experimental Scientists). By integrating material design driven by multimodal models (which combine textual knowledge, chemical compositions, and microstructural information) with high-throughput automated experiments, this platform significantly enhances the speed and quality of catalyst research and development.

One of the core challenges in experimental materials science is how to achieve efficient optimization within a vast chemical design space.

Traditional discovery methods often rely on single-modal active learning frameworks, which utilize only one type of data, such as the mapping relationship between elemental compositions and properties.

The CRESt platform released by the team led by Ju Li at MIT is dedicated to collecting various forms of data through automated experiments and incorporating them into a unified active learning framework, namely Knowledge-Assisted Bayesian Optimization (KABO).

Paper link: https://www.nature.com/articles/s41586-025-09640-5

The robotic system ensures precise control of chemical compositions. High-throughput scanning electron microscopes provide microstructural images, which are then analyzed using computer vision. Meanwhile, large language models embed literature knowledge into the search space.

By vectorizing these different data sources and retaining most of the variance information through principal component analysis, the optimization process becomes more efficient.

The optimized formulations are then mapped back to elemental ratios and tested experimentally, thus forming a closed loop among material design, preparation, and testing.

In addition, another algorithmic innovation is the proposed Bayesian Optimization with Policy-Improved Constraints (BOPIC), which introduces Lagrange multipliers to dynamically adjust the balance between exploration and exploitation, thereby avoiding manual parameter tuning.

Within just three months, CRESt completed over 900 catalyst chemical compositions and more than 3,500 electrochemical tests. It also discovered chemical formulations in both ternary and octonary systems that significantly outperform the traditional optimal pure palladium-based catalysts.

Beyond algorithmic improvements, the researchers also addressed one of the most common issues in experimental science: the lack of reproducibility of experimental results.

Despite the use of robots, inconsistencies in synthesis and testing initially led to significant data noise, reducing the effectiveness of active learning.

To solve this problem, the team collected (through photos and videos) and discussed non-reproducible phenomena in experiments over a long period, such as thermal, electrical, and magnetic effects, as well as experimental errors caused by deeply ingrained human concepts. CRESt employs Vision-Language Models (VLMs) to assist in experiments, diagnose the sources of non-reproducibility, and propose corrective measures.

For example, the model found that a misalignment of the pipette tip on the micrometer scale could cause the carbon paper substrate to shift, resulting in significant deviations in the positions of all samples. In another case, the VLM identified carbonization marks on the surface of a laser-cut wooden sample holder, which led to dimensional changes.

Based on the feedback, the researchers switched to stainless steel holders, significantly improving stability and consistency. To systematically evaluate this diagnostic ability, the research team compiled a small dataset of question-answer pairs based on experimental failure cases.

These examples demonstrate that VLMs can not only diagnose hidden errors but also express them in a language easily understandable by scientists, thereby accelerating the error correction process and reducing manual intervention.

During the experiments, the team discovered an octonary high-entropy alloy catalyst composed of Pd, Pt, Cu, Au, Ir, Ce, Nb, and Cr.

Compared with the pure palladium reference sample, the power density per unit cost of this catalyst increased by 9.3 times, and it achieved the highest performance to date in direct formate fuel cells, requiring only one-fourth of the previous noble metal loading.

X-ray diffraction and Rietveld refinement analysis confirmed that the optimized formulation maintained a single face-centered cubic phase, indicating that the alloy strategy can adjust the local coordination environment while maintaining crystallographic stability.

To understand the mechanism behind the performance improvement, the researchers combined in-situ X-ray absorption spectroscopy (XAS) with density functional theory (DFT) calculations.

The spectroscopic results showed that palladium and platinum remained in the metallic state under reaction conditions, which is crucial because their oxides have almost no catalytic activity.

The doping of Nb, Cr, and Ce introduced subtle structural perturbations without causing significant lattice distortion, thus changing the electronic interactions while maintaining structural integrity.

The DFT calculation results indicated that the energy barrier of the rate-determining step for the palladium sites in the high-entropy alloy along the indirect oxidation pathway was -0.005 eV, compared to 0.706 eV for pure palladium, suggesting a significant improvement in the resistance to carbon monoxide poisoning.

Projected density of states (PDOS) analysis further revealed that the d-band center of palladium in the high-entropy alloy was significantly shifted downward relative to that of pure palladium, thereby weakening the adsorption strength of hydrogen and carbon monoxide and promoting the desorption process.

These theoretical predictions were verified by isotope labeling and CO stripping experiments, which confirmed the higher tolerance to surface poisoning.

The development of CRESt shows that integrating multimodal artificial intelligence with an automated robotic platform can make the exploration of a vast chemical design space, which was previously infeasible, a practical reality. By embedding prior knowledge, achieving adaptive optimization, and using Vision-Language Models to diagnose experimental anomalies, this platform provides a scalable blueprint for accelerating discoveries in chemistry and materials science.

References

https://news.mit.edu/2025/ai-system-learns-many-types-scientific-information-and-runs-experiments-discovering-new-materials-0925

This article is from the WeChat official account "New Intelligence Yuan". Author: LRST. Republished by 36Kr with permission.