Realising high-performance 2D perovskite nanoparticles for efficient light-emitting devices with machine-learning driven experimentation

在过去的十年中，杂化有机-无机钙钛矿（HOIPs）由于其在溶液处理材料中具有特殊的半导体特性，如锐利的能带边缘，高发光率和远程电荷输运，已经得到了大量的研究。金属卤化物钙钛矿由有机分子一价阳离子(A)、金属(B)和卤化物(X)组成，化学计量学为ABX3。阳离子的变化通过减少非辐射损失来提高发光产率。它们的2D钙钛矿类似物的变体-胶体纳米颗粒（NPs）钙钛矿，例如纳米晶体或2D纳米片-有望提高稳定性，高量子产率，在led中也可以在低激发密度下使用，并且具有1D/2D约束的强激子约束。该项目旨在结合材料和实验数据的机器学习（ML）和从头算方法，以及钙钛矿合成的进展，对钙钛矿的光电应用进行全面研究。博士项目的第一个分支是实施一种新的机器学习模型，该模型考虑到当前钙钛矿面临的固有不确定性和可重复性问题。其关键思想是从少量的实验数据出发，构建一个粗糙的贝叶斯机器学习模型；然后，该模型根据探索未知成分空间和开发可能最优成分之间的平衡，建议合成和测试哪些成分，并将实验结果反馈到模型中，以便模型建议要探索的新成分。这种迭代的主动学习方法通过使搜索远离可能是死胡同的组合空间来避免组合搜索。贝叶斯优化已经在数学文献中提出，但在材料科学中的应用迄今为止是有限的。该模型已被证明是成功的一组玩具问题，并接受了机器学习会议研讨会进一步讨论。模型的真正基准测试将发生在从合作者那里收集的真实实验数据和从文献中收集的数据上。博士项目的第二个分支旨在对3D和2D钙钛矿系统的光物理进行全面的从头算研究。众所周知，钙钛矿具有许多不同的稳定多晶相的相复杂性。这些多晶态可以存在于不同的温度和压力范围内。我们的目标是研究钙钛矿系统的结构和稳定性：在二维或三维系统中的相行为是什么？究竟是什么影响了这些系统的相稳定性？缺陷的形成和浓度如何影响相的形成？界面自由能在HOIP体系的末相环境中起作用吗？本文旨在利用密度泛函理论（DFT）和ML力场两种方法来研究这些问题。首先，使用DFT研究各种不同的钙钛矿环境将产生高精度的能量和力场。为了解决缺陷形成的问题，仅用DFT在室温下进行模拟是非常昂贵的。使用ML力场，可以从计算成本降低的数量级中获益，并且具有与DFT相似的精度。通过创建大量的单元细胞，其中唯一的变化是有机阳离子的有序程度，并对这些细胞进行DFT模拟，范围从完全有序到完全无序，研究旨在得出这些有序和无序单元细胞之间的能量比较，以及是否可以最终表明这种有序可以影响系统的带隙。如果确实如此，那么未来的实验工作将不得不警惕这些排序效应，并在合成过程中控制它，以允许另一个微调带隙。

Hybrid organic-inorganic perovskites (HOIPs) have been heavily studied, over the past decade, due to their exceptional semiconducting properties for a solution-processed material, such as sharp band edges, high luminescence yields and long-range charge transport. Metal-halide perovskites are composed of an organic molecule monovalent cation (A), a metal (B) and a halide (X) in the stoichiometry, ABX3. Variations of the cation improve luminescence yields by reductions of non-radiative losses. Variants of their 2D perovskite analogues - colloidal nanoparticle (NPs) perovskites, for example nanocrystals or 2D nanoplatelets - promise increased stability, high quantum yields, also at low excitation densities in LEDs, and strong excitonic confinement from 1D/2D confinement. This projects aims to combine machine learning (ML) and ab initio methods for materials and experimental data, with advances in synthesis of perovskites to perform a comprehensive investigation into perovskites for optoelectronic applications.The first branch of the PhD project is to implement a novel machine learning model which takes into account the inherent uncertainty and reproducibility issues that current perovskites face. The key idea is to build a coarse Bayesian machine learning model starting from a small amount of experimental data; the model then suggests which compositions to synthesize and test based on balancing between exploring unknown composition space and exploiting compositions that are likely to be optimal, and the results from the experiments are fed back into the model for the model to suggest new compositions to explore. This iterative active learning methodology avoids combinatorial searching by biasing the search away from composition space that is likely to be a dead end. Bayesian optimisation has been proposed in the mathematics literature but application in materials science is thus far limited. This model has been shown to be successful for a set of toy problems, and accepted to machine learning conference workshops for further discussion. The true benchmarking of the model will occur with real experimental data, collected from collaborators, and data collected from literature. The second branch of the PhD project is to intended to be a comprehensive ab initio investigation into the photophysics of 3D & 2D perovskite systems. It is known that perovskites exhibit phase complexity with many different stable polymorphs. These polymorphs can exist under a different range of temperatures and pressures. We aim to investigate the structure and stability of perovskite systems: What is the phase behaviour in a 2D or 3D system? What exactly affects phase stability in these systems? How does defect formation and concentration affect phase formation? Does the interfacial free energy play a role in the final phase environment of a HOIP system? It aims to investigate these questions using two methods, density functional theory (DFT) and ML force fields. First, using DFT investigations into a variety of different perovskite environments will yield high-accuracy energies and force fields.To tackle the question of defect formation, this would be very costly to simulate at room temperature with DFT alone. Using ML force fields, one can benefit from orders of magnitude cheaper computational cost, and similar accuracy to DFT. By creating a large number of unit cells, in which the only variation is the degree of order in the organic cations, and conducting DFT simulations on these cells ranging from perfectly ordered to complete disorder, the investigation aims to derive a comparison of energies between these ordered and disordered unit cells and whether it can be shown conclusively that this ordering can affect the bandgap of the system. If it does, then future experimental work would have to be vigilant of these ordering effects and control it during synthesis to allow for another fine-tuning knob of the bandgap.