image: The diagram illustrates the interplay among data acquisition, machine learning, and experiment synthesis. Physical models such as thermodynamics and kinetics can be integrated into ML models as expert knowledge, effectively improving model performance and interpretability.
Credit: ©Science China Press
Advanced functional materials now constitute an extremely active and breakthrough research area, in which the experimental synthesis is one of the most urgent and crucial challenges. Given the multitude of conditions that must be optimized in synthesis routes, chemical synthesis remains a complex and multidimensional challenge. In practice, chemists in typical laboratory settings can only evaluate a limited subset of experimental conditions, relying on chemical literature, experience, empirical data, and simple heuristics to identify the most influential dimensions for reaction success. Fortunately, the emergence of machine learning (ML) techniques brings exciting hope to this dilemma. ML techniques can bypass the time-consuming experimental synthesis and excavate the structure-property relationships, possessing the potential to identify materials with high synthesis feasibility and suggest suitable experimental conditions for chemical reactions. However, issues such as data scarcity and class imbalance frequently hinder the utilization of ML techniques in inorganic materials synthesis.
This review discusses how computational guidelines and ML techniques can accelerate and optimize inorganic material synthesis. Starting from the energy landscape, it introduces physical models and descriptors derived from thermodynamics and kinetics to provide guidance for material synthesis. By embedding the interplay between thermodynamics and kinetics as domain-specific knowledge, both predictive performance and interpretability of ML models are markedly enhanced. The insights presented aim to deepen scientists’ understanding of the thermodynamics and kinetics inherent in the synthesis process, ultimately optimizing experimental design, increasing synthesis efficiency, and paving the way for the development of physics-inspired ML models.
As a powerful tool in the field of material science, ML techniques hold the potential to uncover the structure/process-synthesis relationship. This review offers a comprehensive discussion on ML-assisted inorganic material synthesis, covering aspects such as data acquisition, common material descriptors, ML techniques, and their applications. Primary data acquisition approaches include high-throughput experimental data collection and scientific literature knowledge mining, and applications of ML-assisted inorganic material synthesis are categorized according to different data sources. These innovative efforts establish a closed-loop optimization framework to create an intelligent research paradigm in inorganic material synthesis, significantly increasing the success rate of experiments.
Despite this promising outlook, the use of ML techniques in inorganic material synthesis remains a nascent and evolving field. Even the most state-of-the-art ML models are still unable to provide the accurate predictions regarding the optimal synthesis routes and outcomes. This review also discusses challenges and broader implications in this area. To bridge the gap between computation-guided/ML-assisted strategy and experiments, both theorists and experimentalists are required to contribute their respective expertise and efforts towards this objective. From the theoretical perspective, using the "bottom-up" strategy to construct mathematical models from the atomistic level for complex chemical synthesis processes can facilitate the deep understanding of thermodynamics and kinetics. In addition, material descriptors based on thermodynamics and kinetics can be integrated into ML models, which can improve the performance and interpretability. From the experimental perspective, development of high-quality experimental dataset is the prerequisite of seeking a global phenomenological description of similar processes (or their extended combinations), which can guide further atomic simulations and chemical synthesis. This review would help scientist effectively utilized ML techniques in predicting the outcomes of synthesis experiments and identifying the optimal experimental condition, ultimately speeding up the material discovery cycle.