ChatGPT Research Group for Optimizing the Crystallinity of MOFs and COFs

Abstract

We leveraged the power of ChatGPT and Bayesian optimization in the development of a multi-AI-driven system, backed by seven large language model-based assistants and equipped with machine learning algorithms, that seamlessly orchestrates a multitude of research aspects in a chemistry laboratory (termed the ChatGPT Research Group). Our approach accelerated the discovery of optimal microwave synthesis conditions, enhancing the crystallinity of MOF-321, MOF-322, and COF-323 and achieving the desired porosity and water capacity. In this system, human researchers gained assistance from these diverse AI collaborators, each with a unique role within the laboratory environment, spanning strategy planning, literature search, coding, robotic operation, labware design, safety inspection, and data analysis. Such a comprehensive approach enables a single researcher working in concert with AI to achieve productivity levels analogous to those of an entire traditional scientific team. Furthermore, by reducing human biases in screening experimental conditions and deftly balancing the exploration and exploitation of synthesis parameters, our Bayesian search approach precisely zeroed in on optimal synthesis conditions from a pool of 6 million within a significantly shortened time scale. This work serves as a compelling proof of concept for an AI-driven revolution in the chemistry laboratory, painting a future where AI becomes an efficient collaborator, liberating us from routine tasks to focus on pushing the boundaries of innovation.

Figure 1

Microwave-assisted green synthesis of the crystalline compounds MOF-321 (MOF-LA2-1), MOF-322, and COF-323. (a) Comparison of the framework structures of rod MOFs, MOF-321 (left), and MOF-322 (right), highlighting their distinct organic linkers and aluminum rod SBU, which influence their optimal synthesis conditions. Color code: Al, blue octahedron; C, gray; N, green; S, yellow; O, pink. Hydrogen atoms are omitted for clarity. (b) Chemical structure of COF-323, [Tp₂(DAPy)₃]_{β-ketonenamine}, formed by reticulating 1,3,5-triformylphloroglucinol (Tp) and 2,5-diaminopyridine (DAPy).

Figure 2

ChatGPT research group. (a) Assigned roles of seven ChatGPT-based assistants, each collaborating to assist human researchers and contributing to diverse research tasks at different stages of the synthesis optimization. (b) Flowchart outlining the closed-loop Bayesian optimization process. Each iteration involves three proposed experiments, their execution, data analysis, and integration of the new data into the existing data set to update the surrogate model, upon which the acquisition function is optimized to suggest the next three experiments.

Figure 3

Outcomes of the AI-guided exploration for MOF-321 synthesis. (a) Plot displaying the crystallinity achieved per experiment across a total of 120 reactions, summing to 6,235 min, which is approximately 4.5 days with each experiment lasting 52 min on average. The initial 12 experiments utilized randomly selected conditions, while the subsequent 108 experiments were conducted across 36 iterations, with each iteration comprising 3 experiments. The running average of the crystallinity index, calculated over windows of 3 iterations (9 experiments), is displayed as a pink line. (b) PXRD patterns obtained from representative experimental samples and (c) detailed synthesis parameter distribution for these selected experiments displayed via a radar plot, revealing that the Bayesian search initially covers a broad variety space, later narrowing for fine-tuning. (d) Bar plot illustrating the mean and standard deviation of the crystallinity index for initial experiments (iteration 0) and subsequent iterations grouped into quartiles (iterations 1–9, 10–18, 19–27, and 28–36). The experiments suggested by the BO process significantly improve the average crystallinity compared to the initial 12 random experiments, and an increase in iteration numbers leads to better performance in later iterations. (e) Five scatter plots displaying the evolution of each synthesis parameter suggested by the BO algorithm as a function of iteration number.

Figure 4

Two-dimensional t-SNE dimension reduction scatter plot representing 120 distinct synthesis conditions for MOF-321 (blue) and MOF-322 (red). Prior to reduction, the synthesis parameters (amount of metal, amount of modulator, solvent volume, reaction time, and temperature) are normalized. The color intensity indicates the crystallinity index, with deeper shades signifying higher values. Labels are provided for five representative synthesis conditions from various regions of the scatter plot, illustrating the distinctiveness of certain conditions and the successful identification of multiple conditions with high crystallinity by the BO process. The plot distinctly indicates that the optimal conditions for MOF-321 and MOF-322 differ.

Figure 5

Overlay of gas adsorption–desorption isotherms of MOF-321 and MOF-322, prepared under varying synthesis conditions with different CI values, showing the evolution of optimal synthesis conditions within the search space. (a) Nitrogen sorption isotherms for MOF-321 samples obtained at 77 K. (b) Water vapor sorption isotherms for MOF-321 samples measured at 298 K, demonstrating different sorption capacities. (c) Nitrogen sorption isotherms for MOF-322 samples obtained at 77 K. (d) Water vapor sorption isotherms for MOF-322 samples measured at 298 K, showcasing different sorption capacities. Each panel presents data for six distinct samples of each MOF, underscoring the impact of synthesis conditions on the crystallinity and consequent gas adsorption properties of these MOFs. P, nitrogen or water vapor pressure; P₀, 1 atm; and P_sat, saturation water vapor pressure. Symbols of filled circles denote the adsorption branch, while empty circles denote the desorption branch.