Data-driven Exploration of Metastable Molybdenum Chalcogenides

本研究的目的首先是了解有用材料的性质，其次是通过计算机模拟预测有用材料的性质。一个例子是储存和运输锂的材料的容量，这决定了由它制成的电池的寿命和最大电力。具体来说，这项研究将开发机器学习（ML）方法来进行材料建模，这属于EPSRC人工智能制造未来领域。机器学习是计算机算法的科学，它会随着经验而改进，而不是人类程序员的明确干预。现代化学系统的计算机模拟可以非常精确，无论是预测有用的性质，如锂的容量，还是解开复杂晶体的结构。然而，模拟通常涉及求解（近似）量子力学方程，对于超过大约1000个原子大小的模型来说，这需要非常大的计算能力。我们开发了从量子力学数据中“学习”的方法，在模拟中达到同样的精度水平，但计算时间要少得多。这为模拟中的长度尺度提供了更接近真实设备情况的途径，为我们合理设计和发现更好的材料提供了新的见解。该项目的主要主题之一是训练数据库的构建：一组示例化合物，ML模型可以从中泛化以预测新化合物的性质。我们将研究以下悬而未决的问题：-充分的学习需要什么样的数据？向模型展示大量无序的液体原子足以使其归纳到原子可能存在的所有可能情况吗？或者，例如固体晶体也很重要吗？-训练数据通常包括成千上万个不同环境中的原子示例。我们如何系统地评估它们的信息内容？这很重要，因为计算机内存的限制限制了可以直接用于学习的数据点的数量，所以我们必须选择一个子集——最好是信息量最大的子集。我们怎么知道ML模型有多精确？将模型预测的原子作用力与训练数据进行比较是很简单的，这为我们提供了一种测量模型误差的方法。然而，这种误差似乎与预测大块材料的性质（例如，它的导热性能）不太相关。我们可能会对这样的性质感兴趣，从而设计出有用的材料，因此，我们对为它们计算的值有信心是很重要的。作为ML方法开发的测试案例，我们还将研究元素钼和硫的化合物。选择这些化合物的原因有两个。首先，它们在化学工业中被广泛用作润滑剂，从原油中去除硫，并捕获汞烟雾以阻止它们逸出到环境中。最近，锂被发现可以在其分层结构的薄片之间自由移动，使其应用于电池材料。硫化钼的多种用途使得理解和预测其结构和性质之间的关系变得非常重要。其次，自1975年以来，科学家们一直在争论MoS3的真实结构：钼原子是呈三角形排列还是呈长链排列。由于这些单元形成的无序网络的复杂性，现有的实验证据是开放的解释。迄今为止，计算研究在理解实验结果方面只能提供有限的帮助，因为它们不能使用具有足够多原子的模型来很好地符合现实。我们希望新的机器学习方法能够解决这个难题，在这个过程中证明它们的准确性和实用性。

The goal of this research is first to understand and second predict the properties of useful materials by means of computer simulations. An example is the capacityof a material to store and transport lithium, which determines the life and maximum electrical power of a battery made from it.Specifically, this research will develop Machine Learning (ML) approaches to materials modelling, which falls under the EPSRC area of manufacturing the future withartificial intelligence. ML is the science of computer algorithms that improve with experience, not the explicit intervention of a human programmer. Modern computer simulations of chemical systems can be highly accurate, whether in predicting useful properties such as Lithium capacity or unravelling the structure of complex crystals. However, the simulations usually involve solving (approximately) the equations of quantum mechanics, which requires prohibitively large computing power for models that exceed roughly 1000 atoms in size. We develop methods that 'learn' from quantum-mechanical data to reach that same level of accuracy in simulations, but with far less computing time. This provides access to length-scales in simulations that correspond much more closely to the situation in real devices, giving us new insights to rationally design and discover better materials.One of the main themes of the project is the construction of the training database: a set of example chemical compounds that the ML models generalise from to predict the properties of new compounds. We will examine the following outstanding questions:- What kind of data is required for sufficient learning? Is it enough to show the model lots of disordered liquid atoms for it to generalise to all the possible situations an atom might find itself in, or are e.g. solid crystals important too?- Training data typically includes hundreds of thousands of examples of atoms in different environments. How do we evaluate their information-content systematically? This is important because computer memory limitations restrict the number of data points that can be used directly for learning, so we must select a subset-ideally the most informative ones.- How do we tell how accurate the ML model is? It is simple to compare the force on an atom predicted by the model versus the training data, which gives us one measure of the error of the model. However, this error does not seem to correlate well with that of predictions of properties of the bulk material, e.g. how well it conducts heat. We likely to be interested in such properties to design a useful material, so it is important that we have confidence in the values we calculate for them.As a test case for the ML methods development, we will also study the compounds of the elements Molybdenum and Sulfur. The reasons for choosing these compounds in particular are twofold. Firstly, they are used extensively in the chemical industry as a lubricant, to remove the sulfur from crude oil, and to capture Mercury fumes to stop them escaping into the environment. More recently, Lithium has been found to move freely between the sheets of their layered structure, enabling applications in battery materials. The many uses of Molybdenum Sulfides alone makes understanding and predicting the relationship between their structure and properties important.Secondly, there has been debate among scientists since 1975 over the true structure of MoS3: whether the Molybdenum atoms arrange themselves in triangles or in long chains. Because of the complexity of the disordered network formed by these units, the available experimental evidence is open to interpretation. Computational studies have been able to provide only limited assistance to date in understanding the experimental results because they cannot use models with enough atoms to correspond well to reality. We hope that the new ML methods will be able to resolve the conundrum, in the process proving both their accuracy and utility.