An all-round AI-Chemist with a scientific mind

Abstract

The realization of automated chemical experiments by robots unveiled the prelude to an artificial intelligence (AI) laboratory. Several AI-based systems or robots with specific chemical skills have been demonstrated, but conducting all-round scientific research remains challenging. Here, we present an all-round AI-Chemist equipped with scientific data intelligence that is capable of performing basic tasks generally required in chemical research. Based on a service platform, the AI-Chemist is able to automatically read the literatures from a cloud database and propose experimental plans accordingly. It can control a mobile robot in-house or online to automatically execute the complete experimental process on 14 workstations, including synthesis, characterization and performance tests. The experimental data can be simultaneously analysed by the computational brain of the AI-Chemist through machine learning and Bayesian optimization, allowing a new hypothesis for the next iteration to be proposed. The competence of the AI-Chemist has been scrutinized by three different chemical tasks. In the future, the more advanced all-round AI-Chemists equipped with scientific data intelligence may cause changes to the landscape of the chemical laboratory.

Keywords: AI-Chemist; all-round research; computational brain; machine reading; mobile robot.

Figure 1.

Design of the all-round AI-Chemist with a scientific mind. (A) Three modules of the AI-Chemist: a machine-reading module, a mobile robot module and a computational brain module. (B) The workflow of the AI-Chemist and the functions of each module.

Figure 2.

The service platform with web browser-based HMI. (A) Machine-reading module for capturing existing knowledge and transferring to machine-understandable structured data by NLP. (B) User-friendly GUI to monitor the status of chemical workstations and robots in real time. (C) Experimental workflow diagram: from scientific hypotheses to workstation status, then to experimental design and template. (D) Experimental data such as electrochemical active surface area and current–voltage curve collected and displayed on the service platform. (E) Experimental archives for building the experimental database and training the prediction model.

Figure 3.

Intelligent chemical laboratory and mobile robot. The intelligent chemical laboratory with capabilities of robotic path planning, robotic control and connection, and smart chemical operations and the mobile robot equipped with control system, dual-lidar-based localization system, six-degree-of-freedom robotic arm and large scalable loading platform.

Figure 4.

Chemical experiments performed by mobile robot and workstations. (A) Peak fluorescence intensity with increased concentrations of berberine chloride (BBR). (B) Fluorescence images of the solution in (A) under UV light (365 nm). (C) PL spectra of BBR chloride in solution with different concentrations of BBR chloride. Excitation wavelength: 405 nm. (D) Tunable hydrogenation of MoO₃ by Cu–acid treatment with simulated crystal structure of H_xMoO₃ with different concentrations of H-dopants in the lattice. (E) The UV–Vis spectra of the RhB solution after photocatalytic degradation. (F) Characterization of the photocatalytic degradation efficiency of the RhB (left column, reaction conditions: [cat] = 0.1 g), solution volume: 10 mL H₂O, [RhB] = 10⁻⁵ mol/L) and the yield of dye-sensitized photocatalytic H₂ production (right column, reaction conditions: [Eosin-Y] = 0.6 mM, [TEOA] = 1.5 M, [cat] = 0.1 g, solution volume: 10 mL. All the light sources are 25 W white LED lamps). The error bars show the systematic average error and the comparison with the manual work result (blue error bar). (G) Schematic diagrams of experiment flow for dye-sensitized photocatalytic water-splitting experiment without (top) and with (bottom) the multi-task dynamic optimization. (H) Schematic diagrams of experiment flow for electrocatalytic oxygen evolution reaction experiments without (top) and with (bottom) the multi-task dynamic optimization.

Figure 5.

An all-round chemical research conducted by the AI-Chemist. (A) The order of metal recommendation and the frequency of metal co-occurrence by machine reading. (B) The overpotential values of 207 Try–Error experiments carried out by the mobile robot and workstations. (C) An example of the simulated structure generated by molecular dynamics. (D) The simulated OER reaction path. (E) The prediction results of three catalytic properties by neural network models, where G_OH is the free energy change of the hydroxyl adsorption, G_O is the free energy change of the oxygen atom adsorption and Δe is the charge transfer during hydroxyl adsorption. (F) The prediction results of overpotential by the neural network model calibrated by experimental data. Inset: dimensionality reduction plot by principal component analysis (PCA) for predicted overpotentials of all exhaustive samples. (G) Kiviat diagram of composition ratios and (H) polarization curves of the optimal sample suggested by the Bayesian model and the best samples by Try–Error experiments.