Revolutionizing Chemistry and Material Innovation: An Iterative Theoretical-Experimental Paradigm Leveraged by Robotic AI Chemists

Open AccessCCS ChemistryMINI REVIEWS28 Nov 2024Revolutionizing Chemistry and Material Innovation: An Iterative Theoretical-Experimental Paradigm Leveraged by Robotic AI Chemists Baicheng Zhang, Zhuoying Zhu, Huirong Li, Jiaqi Cao and Jun Jiang Baicheng Zhang , Zhuoying Zhu *Corresponding authors: E-mail Address: [email protected] E-mail Address: [email protected] , Huirong Li , Jiaqi Cao and Jun Jiang *Corresponding authors: E-mail Address: [email protected] E-mail Address: [email protected] https://doi.org/10.31635/ccschem.024.202404860 SectionsAboutAbstractPDF ToolsAdd to favoritesDownload CitationsTrack Citations ShareFacebookXLinked InEmail Chemistry and material innovation is undergoing a transformative shift with the integration of advanced computational and experimental technologies over the past few decades. More recently, the advent of automated workflows, machine learning (ML) techniques, and robotic experiments have elevated scientific research to unprecedented levels. We have explored the iterative theoretical-experimental paradigm, leveraged by robotic artificial intelligence (AI) chemists to bridge the gap between high-volume theoretical data and high-dimensional experimental data in this review. By combining automated high-throughput computations, ML models, and robotic large-scale experiments, this novel protocol aimed to accelerate data-driven chemistry innovation and materials discovery. Successful applications achieved by this paradigm include nanomaterials, high-entropy alloy catalysts, optical thin films, and oxygen evolution reaction (OER) catalysts from Martian meteorites. We have highlighted the potential for this paradigmatic evolution to redefine research methodologies and promote the next generation of precise and intelligent chemistry innovation. Download figure Download PowerPoint Introduction Over the centuries, chemistry has predominantly relied on a trial-and-error approach, making it a lengthy and unpredictable procedure.1,2 High-performance computing (HPC) began to significantly impact high-throughput quantum chemistry simulations and accelerate data-driven materials discovery especially when the Materials Genome Initiative (a U.S. federal effort aimed at accelerating advanced materials development, providing policy, resources, and infrastructure to reduce time and cost compared to traditional methods) was launched in 2011.3–5 Machine learning (ML) in chemistry and materials science has exponentially increased from a small niche to an entire subfield in the past few years.6–9 For instance, a recent work by Google Deepmind used graph networks to identify 2.2 million hypothetical materials, representing an order-of-magnitude expansion in stable materials found to date though further experimental validations are necessary.10,11 With the advancements in algorithm developments, computing power, and robotic techniques, researchers are ushering in an unprecedented era in chemistry innovation and materials discovery.12–14 The gap between theoretical calculations and real-world phenomena objectively exists. Theoretical chemistry is competent in providing low-dimensional solutions based on simplified or idealized assumptions. Many quantum chemistry simulations investigate the intrinsic reactions, which are valuable for giving general trends and facilitating high-throughput calculations but fail to calibrate observable phenomena under complex circumstances. On the other hand, experimental data reveal high-dimensional information about matters but are typically limited in volume, which can result in local optima rather than comprehensive global solutions. A combined approach that leverages the strengths of both theoretical simulations and experimental data have proven to be promising in fostering data-driven chemistry innovation and accelerating materials discovery.14,15 Previous data-driven studies have largely adhered to a linear workflow, progressing from simulations to experiments through isolated operational steps.16–19 This one-way sequence approach lacks a dynamic connection between theory and practice. To address the misalignment between simulation results and experimental outcomes, an iterative computational-experimental paradigm, empowered by robotic artificial intelligence (AI)-chemists have emerged and demonstrated its capabilities in various scientific scenarios.12–15,20–22 Equipped by powerful tools of natural language processing (NLP),23–26 ML,7,27,28 theoretical chemistry and robotic experimentation.12–14 AI chemists can achieve research tasks such as intelligent literature reading, high-throughput computing, and automated all-round experiments with a high degree of autonomy. We believe that the advantages of this new paradigm at least lie in the following aspects: 1. Bridge the gap between low dimensional, large volume theoretical data and high dimensional, small volume experimental data. 2. Help to discover new knowledge by running counter to researchers' chemical intuition. 3. Preserve particularly functional compositions that are not easily accessible, e.g., metastable materials.29–31 In this mini-review, we first introduce recent theoretical and experimental advances in accelerating chemical research. Then we summarized the accomplishments of robotic AI chemists on materials design in nanomaterials, high-entropy alloy (HEA) catalysts, and optical film. Ultimately, we provided an outlook on future developments that could promote the next generation of precise and intelligent chemistry innovation. We hope this review summarizes the state-of-the-art advances and progress in the field, and elucidates the blueprint for a novel chemical research paradigm that seamlessly integrates theoretical and experimental approaches leveraged by robotic AI-chemists (Figure 1a,b). Figure 1 | Comparison between theoretical or experimental approaches and an iterative theoretical-experimental paradigm in research highlights their respective strengths and limitations. (a) Theoretical chemistry generates large volumes of data but with low-dimensional outputs, limited by simplified assumptions, leading to inevitable deviations. Experimental chemistry produces high-dimensional, real-world results but yields low data volume, constrained by locally optimal solutions and (b) Schematic of an iterative theoretical-experimental paradigm designed for global optimization through continuous feedback. Download figure Download PowerPoint Advanced Tools for Iterative Theoretical-Experimental Chemistry and Material Innovation In the iterative theoretical-experimental paradigm, first-principles computations generate a substantial amount of theoretical data, offering a rough but comprehensive global perspective. Robotic experiments contribute high-dimensional empirical data for calibration, enhancing model accuracy. ML then integrates both theoretical and experimental data, constructing predictive models that refine computational sampling and guide subsequent experiments. Through iterative cycles of refinement, the model evolves to capture both global trends and high-dimensional details, ultimately leading to the discovery of the global optimal solution. Automated workflow for high-throughput computations The fundamental laws governing electron behaviors enable the predictions of chemical properties. Theoretical methods range from Hartree–Fock (HF),32 density functional theory (DFT)33 to Quantum Monte Carlo (QMC),34 and apply to various targeted systems, from small molecules to periodic crystals. With advancements in computing power, high-throughput computations offer great advantages, including (1) generating high-volume data faster and cheaper than experiments; (2) providing theoretical references and guiding efficient experiments; and (3) collecting clean and consistent datasets for machine-learning training. DFT utilizes electron density as the kernel variable and employs Kohn-Sham equations along with exchange-correlation functionals to approximate the energy. Its strength lies in the excellent balance between accuracy and computational cost, making it extensively used in high-throughput computations. Continuous efforts on functionals35–38 have been made to address the limitations of existing functionals and to provide a wide range of choices for different simulation systems with optimized accuracy, applicability, and reliability. DFT places significant demands on HPC for in-silico materials design and leverages the infrastructure of large databases to enhance its impact. The three most-cited databases in the material science field (Materials Project,10 Open Quantum Materials Database,39 and AFlow40) are built primarily on DFT-generated data. Analysis packages [e.g., pymatgen,41 atomic simulation environment (ASE),42 and matminer43] and automation workflows (e.g., atomate,44 atomate244) are fruitful products by the demand of high-throughput computations, at the other end of the scale, promoting further chemical and material discovery at large-scale. From the experience in the past decade, we have learned that generically deploying robust software infrastructure is crucial for fully automated high-throughput computations. These standardized workflows streamline the user experience and facilitate large-scale computing under effective management. Consequently, we have witnessed the successful applications of high-throughput computations managed by automated pipelines in (1) spectra simulations (e.g., Raman spectra45 and X-ray absorption spectroscopy46,47); (2) cathode48–51 and solid electrolyte19,52 in Li-ion batteries; (3) lattice thermal conductivity prediction53,54 for thermoelectrics; and (4) energy prediction (e.g., formation energy and absorption energy).55–57 It is worth noting that a previous study58 indicated that a threshold of ∼ 10,000 data points is required to pre-train more accurate high-quality deep learning models. For medium or small datasets, techniques such as transfer learning59 and multifidelity modeling60 may be necessary to reach acceptable accuracy. Efforts in maintaining large databases, developing software libraries, and creating automated workflows have aimed to supply AI-ready data, exponentially accelerating chemistry and material innovation in the new era. Machine learning for chemical discoveries ML and AI are extensively used in chemistry and materials science research, aiding researchers in establishing relationships between structure, properties, and spectra. In these fields, ML is primarily employed for predicting properties, optimizing schemes, and designing reverse generation. Descriptors, as a crucial part of ML for chemistry, provide a numerical representation of molecular and material properties. Here we list four categories: property descriptors, sequence descriptors, structural descriptors, and spectral descriptors and briefly explain how they serve distinct purposes as follows: 1. Property Descriptors: These utilize physical and chemical properties to numerically represent substances like molecules and crystals. Researchers typically select properties with inherent physical relevance to the prediction target such as molecular weight, partition coefficient (LogP), hydrogen bond donors and acceptors, and topological polar surface area. These properties are commonly used for drug molecules.61 In a single-atom electrocatalyst field, the spin moment linked to electronic spin states has proven effective in the intelligent design of electrocatalysts.62 Property descriptors have strong physical relevance, making results interpretable, but selecting suitable descriptors requires expertise and may not capture all subtle interactions in complex systems. 2. Sequence Descriptors: Organic molecules have efficient notation systems to encode their bonding structures into linear strings of characters, naturally serving as sequence descriptors like simplified molecular input line entry system (SMILES), self-referencing embedded strings (SELFIES), and International Chemical Identifier (InChI). These descriptors encode information about atoms, covalent bonds, aromaticity, radicals, formal charges, cis-trans isomerism, and optical isomerism into strings, enabling the use of powerful language models in chemistry due to their compatibility with text-based processing tools. SMILES descriptors have successfully predicted molecular energy, electronic, and thermodynamic properties.63 Transformer and Generative Pretrained Transformer (GPT) models can also generate new drugs and materials based on SMILES and SELFIES.64–68 Sequence descriptors are compatible with NLP models, enabling powerful predictions and molecule generation. However, they can sometimes lack direct interpretability, which makes understanding the physical meaning behind the encoded sequences challenging. 3. Structural Descriptors: These use molecular and material structural features as inputs for ML models, including bond lengths, bond angles, internal coordinates, and molecular topology. For instance, structural descriptors based on bond lengths, angles, dihedral angles, and topology have proven effective in accurately predicting bond dissociation energy.69,70 Structural descriptors are advantageous for capturing the geometrical and topological aspects of molecules, leading to accurate property predictions, but they can be computationally intensive to derive, especially for large or complex systems. 4. Spectral Descriptors: Spectra such as infrared (IR), mass spectrometry (MS), and nuclear magnetic resonance (NMR) are mathematical representations of molecules and materials. Spectra offer both theoretically calculable and experimentally measurable characteristics, enabling the acquisition of extensive theoretical data for ML training and quick transition to experimental conditions. For example, IR and Raman spectra of adsorbates serve as spectral descriptors to predict adsorption energies and charge transfer properties.71 Through chemical reaction predictions, NMR chemical shifts are incorporated as descriptors, helping ML models discern subtle commonalities and differences across various reaction types.72 Spectral descriptors are beneficial as they provide both calculable and measurable data, making them versatile for ML model training and experimental application; however, they can be challenging to interpret directly and require sophisticated analysis to extract meaningful features. Advanced ML techniques enable significant advancements in predicting and designing new materials and molecules in chemistry. By leveraging ML techniques, researchers can uncover intricate patterns and relationships within complex chemical data, leading to more accurate predictions and innovative designs. Supervised learning requires labeled training models and aids in the precise prediction of molecular properties,73–75 chemical reaction yields,76–78 and protein spectra,79,80 which are crucial for understanding chemical behaviors and developing new compounds. Commonly used ML models include linear regression, logistic regression, decision trees, support vector machines (SVMs), random forests, and neural networks. Unsupervised learning helps to discover hidden patterns and structures in large datasets without labeled data, facilitating chemical data clustering,81–83 material property prediction,84 solar cell material research,85 and crystallography characterization86 through various clustering methods (e.g., K-means, hierarchical clustering). Reinforcement learning optimizes experimental strategies, enhancing the efficiency of novel material design86 as well as digital quantum simulation,87 chromatography,88 electrochemistry,89 and crystallography analysis.90 Generative models, such as generative adversarial networks (GANs) and variational autoencoders (VAEs), create new data that meets specific criteria, enabling the inverse design of materials such as semiconductor materials,91 generation of inorganic solid materials,92 and porous materials93 with desired properties. Overall, the integration of ML in chemistry accelerates research, reduces experimental costs, and drives the development of novel materials and molecules with tailored functionalities, significantly advancing the field. Robotic experiments for chemical discoveries To achieve the global optimum predicted from massive ML predictions, it is essential to conduct fast and reproducible experiments. This requires concerted efforts to develop a comprehensive and high-volume database that encompasses synthesis, characterization, and materials processing data, where robotic experiments could play a significant role. Here we introduce the state-of-the-art robotic chemists developed by Zhu et al.14 who utilized automated robotic techniques to leverage high-throughput experiments. The robotic chemist is built on a wheeled mobile robot base, equipped with a 6-axis mechanical arm, which connects more than twenty automated workstations. It can execute a comprehensive and flexible suite of user-designed experimental protocols. These workstations are categorized into three types: (1) automated synthesis, (2) automated sampling, and (3) automated characterization and testing, which mimic the fundamental operations and workflows similar to those in conventional inorganic chemistry laboratories. The whole system is controlled and directed by the robotic instruction set. Among the three categories, the automated synthesis workstations primarily handle the synthesis procedures, including dispensing and mixing of reactants under different reaction conditions (e.g., heating and illumination). The capability and flexibility for various syntheses rely on standard templates with customizable options. For instance, sample mixing is facilitated through three different types of workstations: ultrasonic, magnetic stirring, and shaking beds, allowing researchers to choose based on the specific experimental requirements. In contrast to the A-Lab (an autonomous laboratory integrating robotics, ab initio databases, ML, and data mining),94 which primarily focuses on solid-state reactions, the robotic chemist's automated synthesis is optimized for handling liquid samples, with a specialization in processing solutions and suspensions. Liquid transfer and mixing can be efficiently performed with a pipette, avoiding the residues commonly encountered with solid precursors, which is a universal challenge in any robotic experimental platform. Automated sampling currently comprises sample preparation for electrochemical testing, IR, and X-ray diffraction (XRD) spectral characterization. The operations are facilitated by a track-mounted robotic arm in conjunction with a substrate distribution platform. The robotic arm selects appropriate substrates and precisely dispenses samples based on specific testing requirements. Sample preparation includes solvent drying via methods like heating or IR, chosen at specified power levels. Robotic sampling, akin to manual experimentation, involves numerous variables like substrate material, thickness, liquid volume and concentration, solvent type, drying power, and duration. Careful planning is essential to tailor optimized parameters in each experiment, by balancing the limited operations in the robotic platform and myriad experimental variables. Automated characterization and testing contain a series of spectroscopic workstations and generate a collection of characterization data. When combined with data from the aforementioned synthesis and sampling processes, a standardized set of experimental data is formed. Unlike traditional datasets, these not only include data on experimental outcomes but also contain details of the experimental process, which are seldom highlighted in experimental articles or ML studies. Robotic experiments, due to their high repeatability and standardized operations, are uniquely equipped to overcome this issue, thus facilitating the utilization of the metadata of experimental processes. Systematically screening and optimizing conditions from numerous experimental variables is a labor-intensive task for human researchers. Robotic AI-chemists offer a significant advantage in this regard, as they can tirelessly and accurately conduct a large number of experiments. The extensive experiments achieved by robotic chemists can further refine and calibrate ML models, which is a crucial part of the iterative, theoretical-experimental loop (Figure 2a–d). Figure 2 | Schematic for automated high-throughput computations, ML models, and robotic experiments, illustrating their growing use based on publication trends in the past ten years. (a) The automated computational workflow using DFT to calculate electronic structure; (b) Key components of ML in chemistry: descriptors, methods, and applications; (c) Schemes and the workflow for robotic experiments connected by automated workstations; and (d) Trends in high-throughput computations, ML in chemistry, and robotic experiments, as shown by annual publications (2013–2023) from Web of Science using relevant keywords ("High-throughput Computation/Framework/Screening + Chemistry," "Machine Learning + Chemistry" and "Robotic Experiments"). Download figure Download PowerPoint Applications Achieved by Iterative Theoretical and Experimental Approaches Robotic chemistry labs now integrate advanced techniques such as data mining, theoretical computations, ML, and robotic experimentation, each applied strategically to specific challenges. Examples include optimizing chalcogenide morphology using data mining and robotics, analyzing Martian meteorites with laser-induced breakdown spectroscopy (LIBS) to synthesize catalysts for oxygen production, and investigating HEA electrocatalysts through fully autonomous workflows combining simulations, experiments, and ML. For optical thin films, inverse material design was achieved through ML and robotics, while Bayesian optimization (B-OPT) enhanced synthesis efficiency in organic chemistry. These examples highlight the versatility of robotic systems in solving complex scientific problems. The detailed case studies below demonstrate the impact of automation and ML on experimental outcomes. Data mining for tuning perovskite morphology Colloidal nanocrystals (NCs) hold substantial promise in electrochemical, photochemical, and pharmaceutical applications, with their performance being highly dependent on morphology. However, the synthesis and material characterization in this domain are complex and resource-intensive, even with robotic high-throughput experiments. Zhao et al.15 address this challenge by developing a robotic platform for the synthesis of colloidal NCs emphasizing morphological control (see Figure 3a–d). The study introduces a framework that integrates data-driven automated synthesis with morphology-oriented inverse design. Key synthesis parameters were initially identified through data mining from more than 1300 literature. The platform then executed automated synthesis and in-situ characterization, followed by ex-situ validation. ML models trained on an expanding experimental database identified correlations between the morphologies and structure-directing agents. This approach enabled the controllable synthesis of NCs with desired morphologies, demonstrating its efficiency and effectiveness compared to traditional methods. In this work, data mining methods provide reliable ranges of synthesis conditions (e.g., concentrations), which guide the researchers in designing an experimental protocol and accelerate high-throughput robotic experiments to achieve the desired morphology of gold NCs and double perovskite. Figure 3 | Morphology control of gold and double-perovskite NCs. (a) Data mining from 1300–1400 papers yielded the distribution of synthesis parameters for five raw materials: cetyltrimethylammonium bromide (CTAB), AgNO3, HCl, AA, and HAuCl4. The highest frequency is represented by L2, with L1 and L3 for lower and upper boundaries, respectively. (b) The flowchart combines robotic experiments and ML to achieve optimized morphology (nanosize = 78 nm for double perovskite, aspect ratio = 4.06 for Au NCs) based on data-mined synthesis parameters. (c) The robots autonomously designed and executed high-throughput experiments for double perovskite using mined data. (d) Five high-throughput sets were conducted using orthogonal (O), single-factor (S), double-factor (D), triple-factor (T), and inverse design (I) approaches for gold NC synthesis. Copyright 2024 The American Chemical Society. Download figure Download PowerPoint LIBS enabled on-site production of OER electrocatalyst from Martian meteorites As human civilization advances and the demand for living space and resources expands, Mars has captured the attention and exploration efforts of scientists due to its potential for human habitation, despite its lack of oxygen. Therefore, producing oxygen using local Martian resources is the primary task for Mars immigration. In recent work by Zhu et al.22 AI chemists can synthesize high-performance oxygen evolution reaction (OER) catalysts from Martian meteorites using fully automated processes without human intervention, meeting practical needs for Mars development (see Figure 4). Figure 4 | AI-chemist creates OER catalyst from Martian meteorites. The elemental composition of the Martian meteorite was first analyzed using Libs analysis, and then the metal hydroxides were obtained by acid addition, dissolution, and extraction, alkali addition, solid–liquid separation, waste liquid removal, and drying, and the metal proportions were then varied and the corresponding overpotentials were obtained using electrochemical performance tests. Theoretical datasets of metal compositions and OER catalytic properties were obtained through high-throughput molecular dynamics simulations and first-principles calculations. An ML model for theoretical data was constructed fine-tuned and optimized using the experimental data to enable prediction of overpotentials from experimental formulations. The experimental formulation of the best catalyst was then obtained using a B-OPT algorithm and validated experimentally. Copyright 2023 Spring Nature. Download figure Download PowerPoint AI chemists first use LIBS to analyze the elemental composition of Martian meteorites. Through a series of experimental manipulations, they derived metal hydroxides and conducted 243 experiments by varying metal ratios to obtain corresponding overpotentials. Following this, high-throughput molecular dynamics simulations and first-principles calculations are employed to determine the OER catalytic properties (ΔGOH*, ΔGO*-OH*, Δq) of 29,902 different metal components. Then a theoretical model based on ML is constructed to connect the metal components with their catalytic properties. Subsequently, after incorporating previous experimental data, a trainable output layer connects to the theoretical model and predicts more experimental overpotentials. The final model integrates theoretical and experimental data, achieving the goal of predicting overpotentials from experimental formulations. Ultimately, employing B-OPT algorithms, AI chemists identify the optimal catalyst formulation from over 3 million possible compositions. This catalyst formulation demonstrates the capability to operate for over 550,000 seconds at a current density of 10 mA cm−2 with an overpotential of 445.