D3SC: Machine Learned Free Energies of Compounds

科罗拉多大学博尔德分校的Charles Musgrave和合作者Aaron Holder得到了化学系、化学、生物工程、环境和运输系统系的支持，开发和应用机器学习方法来发现新材料。虽然元素周期表提供了许多可能形成材料的元素组合，但这些化合物中只有一小部分是稳定的或具有特定应用所需的特性。此外，在大量可能的化合物中，只有大约1000种在高温下具有已知的特性。在过去的50年里，计算化学家利用量子力学方程来发现新材料。然而，筛选大量的候选材料用于特定的技术应用仍然过于计算要求是实际的。最近，统计学习方法已经发展起来，它可以从大量数据中提取系统信息，以训练高可靠的“人工智能”模型来预测新系统的属性。在这个项目中，马斯格雷夫教授和霍尔德教授正在使用机器学习方法来预测材料的稳定性、结构和化学反应性。预测的性质可以用来确定催化技术上重要的反应的候选材料，例如将水分解成氧气和氢气，将二氧化碳转化为有用的产物，或者从氮和水中“绿色”合成氨。这些模型可以作为机器学习计算机代码在公共存储库中获得，也可以通过公共访问的数据库获得。该项目旨在培训高中生、本科生和研究生开发和应用最先进的机器学习方法，用于化学和化学工程应用。研究人员参与了科罗拉多大学通过领导力和多样性扩大机会（BOLD）中心。将机器学习和化学的新概念整合到课程中，并通过部门LearnChemE YouTube平台进行整合。该项目结合了电子结构、热力学、计算科学和机器学习方面的专业知识，研究分子最基本的性质之一——吉布斯自由能G(T)。数据驱动的方法利用了结果表明，在定体积准谐波近似中计算的振动熵和亥姆霍兹自由能-这些量对G(T)有重要贡献，但在量子力学计算方面具有计算挑战性-具有系统的温度依赖性，可以使用机器学习准确有效地预测，再加上材料的化学成分知识。研究人员正在扩展这一观察结果，将机器学习方法直接应用于G(T)模型，使用数百个分子的实验数据进行训练和描述符提取。由此产生的描述符被用于预测无机晶体结构数据库中约20,000种独特成分的热化学数据，反过来，用于计算与温度相关的凸壳相图和固态反应平衡。模型和G(T)数据可在大型数据库中获得。新的方法使人们能够发现总的趋势和新的化学知识，了解温度和组成对反应性、合成性、稳定性和亚稳态的影响。除了为分子和反应的热化学提供深入的见解外，这项研究还能够识别异常情况，这些异常情况可能表明系统中的新特性正在改变分子的行为。例如，依赖于温度的涌现现象或量子现象创造了独特的材料特性。尽管先进材料在广泛的技术领域具有重要的技术和经济意义，但人们对其产生稳定性和反应性的详细行为仍知之甚少。新技术和热化学数据库的潜在应用包括使用氧化还原材料的热化学水裂解，通过化学环合成氨，氧化化学，碳热还原氧化物，以及通过分子氢或其他还原剂还原分子。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

Charles Musgrave of the University of Colorado Boulder and collaborator Aaron Holder are supported by the Division of Chemistry, and the Division of Chemical, Bioengineering, Environmental, and Transport Systems, to develop and apply machine learning approaches for the discovery of new materials. While the periodic table provides a many possible combinations of elements from which to form materials, only a fraction of these compounds will be stable or have desirable properties for particular applications. Furthermore, of the large number of possible compounds, only about one thousand have known properties at elevated temperatures. For the past fifty years, computational chemists have used equations of quantum mechanics to discover new materials. However, screening large numbers of candidate materials for a specific technological application remains too computationally demanding to be practical. Recently, statistical learning approaches have been developed which can extract systematic information from large quantities of data to train highly reliable "artificial intelligence" models for predicting properties of a new system. In this project, Professors Musgrave and Holder are using machine learning approaches applied to predict the stabilities, structures, and chemical reactivity of materials. The predicted properties can then be used to identify candidate materials for catalyzing technologically-important reactions, such as splitting water into oxygen and hydrogen, converting carbon dioxide into useful products, or the 'green' synthesis of ammonia from nitrogen and water. The models are available on public repositories as machine learning computer codes, and through publicly-accessible databases. The project is training high school, undergraduate and graduate students in the development and application of state-of-the-art machine learning methods for chemistry and chemical engineering applications. The researchers participation in the Broadening Opportunity through the Leadership and Diversity (BOLD) Center at University of Colorado. The incorporation of new concepts in machine learning and chemistry are integrated into courses and through the departmental LearnChemE YouTube platform.This project combines expertise in electronic structure, thermodynamics, computational science, and machine learning to study one of the most fundamental properties of molecules--the Gibbs free energy, G(T). The data-driven approach takes advantage of results showing that the vibrational entropy and Helmholtz free energy computed in the constant-volume quasiharmonic approximation - quantities that critically contribute to G(T) but are computationally challenging to calculate quantum-mechanically - have systematic temperature dependence and can be accurately and efficiently predicted using machine learning, coupled with knowledge of the chemical composition of the material. The researchers are extending this observation to apply machine learning methods to model G(T) directly, using experimental data for several hundred molecules for training and descriptor extraction. The resulting descriptors are being used to predict thermochemical data for ~20,000 unique compositions tabulated in the Inorganic Crystal Structure Database, and in turn, to compute temperature-dependent convex hull phase diagrams and solid-state reaction equilibria. The models and G(T) data are available on large databases. The new methodology is enabling the discovery of general trends and new chemical knowledge of the effects of temperature and composition on reactivity, synthesizability, stability and metastability. In addition to providing deep insights into the thermochemistry of molecules and reactions, this research is enabling the identification of anomalies that may indicate systems where emerging properties are altering the behavior of the molecule. For example, where temperature-dependent emergent or quantum phenomena create unique materials properties. Despite the technological and economic importance of advanced materials in a broad range of technologies, much is still unknown about the detailed behavior that give rise to their stability and reactivity. Potential applications of the new techniques and thermochemical databases produced include thermochemical water splitting using redox materials, ammonia synthesis by chemical looping, oxidation chemistries, carbothermal reduction of oxides, and reduction of molecules by molecular hydrogen or other reductants.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.